Paper deep dive
CHILL-Harness: Counterfactual Harness Learning for Efficient Reasoning in Long-Horizon Agents
Jiarun Fu, Lizhong Ding, Sida Chen, Honglei Xin, Chunhui Zhang, Pengqi Li, Qiuning Wei, Ye Yuan, Guoren Wang
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%
Last extracted: 8/1/2026, 10:26:50 AM
Summary
The paper introduces CHILL-Harness, a causal framework for adaptive orchestration in long-horizon LLM agents. It addresses the limitation of static harness policies by using Causal Intervention Effect Learning (CIEL) to estimate workflow advantages and Advantage-Realizing Causal Orchestration (ARCO) to selectively apply counterfactual interventions, thereby improving efficiency and task success.
Entities (5)
Relation Signals (5)
CHILL-Harness → contains → CIEL
confidence 95% · we develop Causal Intervention Effect Learning (CIEL) as its effect-estimation component of CHILL-Harness
CHILL-Harness → contains → ARCO
confidence 95% · We further introduce Advantage-Realizing Causal Orchestration (ARCO) as its realization component
CHILL-Harness → improves → execution efficiency
confidence 92% · CHILL-Harness ... improving reasoning and execution efficiency while preserving task performance
CIEL → estimates → workflow advantage
confidence 90% · CIEL ... estimates their context-conditioned intervention-relative advantages from confidence-weighted execution evidence
ARCO → realizes → workflow adaptation
confidence 90% · ARCO ... authorizes workflow adaptations only when they exhibit sufficient causal–operational advantage
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Agent harnesses have become the operational infrastructure of modern large language model agents, coordinating context, tools, verification, and execution control to translate latent model capability into reliable long-horizon behavior. However, reliable long-horizon behavior requires harness control to adapt to task demands, execution environments, and evolving execution states, whereas current harnesses predominantly rely on hand-crafted or globally fixed policies; this mismatch manifests as unnecessary computational overhead and, in adverse cases, reduced task success. To address this limitation, we formulate the task of enabling adaptive orchestration in harness systems as a causal learning problem and propose Counterfactual Harness Intervention Learning for Long-Horizon Agents (CHILL-Harness). CHILL-Harness intervenes at the orchestration layer to enable advantage-guided workflow adaptation, thereby improving reasoning and execution efficiency while preserving task performance. Specifically, we develop causal intervention effect learning as the effect-estimation component of CHILL-Harness to estimate intervention-relative workflow advantage from confidence-weighted execution evidence and identify advantageous workflow adaptations. We further introduce advantage-realizing causal orchestration as its realization component to adaptively allocate counterfactual reasoning and realize only workflow adjustments supported by sufficient expected advantage. Finally, we incorporate a success-preserving objective and advantage-margin authorization constraints into CHILL-Harness to promote reliable adaptation. Extensive experiments on heterogeneous long-horizon tasks spanning information seeking, software engineering, and terminal interaction show that CHILL-Harness consistently preserves or improves task success while substantially reducing token consumption and execution time.
Tags
Links
- Source: https://arxiv.org/abs/2607.25825v1
- Canonical: https://arxiv.org/abs/2607.25825v1
Trouble viewing inline? Open PDF directly →
Full Text
95,512 characters extracted from source content.
Expand or collapse full text
CHILL-Harness: Counterfactual Harness Learning for Efficient Reasoning in Long-Horizon Agents Jiarun Fu1 Lizhong Ding1,∗ Sida Chen1 Honglei Xin1 Chunhui Zhang1 Pengqi Li1 Qiuning Wei1 Ye Yuan1 Guoren Wang1 1School of Computer Science and Technology, Beijing Institute of Technology, Beijing, China Contact: jrfu@bit.edu.cn Abstract Agent harnesses have become the operational infrastructure of modern large language model agents, coordinating context, tools, verification, and execution control to translate latent model capability into reliable long-horizon behavior. However, reliable long-horizon behavior requires harness control to adapt to task demands, execution environments, and evolving execution states, whereas current harnesses predominantly rely on hand-crafted or globally fixed policies; this mismatch manifests as unnecessary computational overhead and, in adverse cases, reduced task success. To address this limitation, we formulate the task of enabling adaptive orchestration in harness systems as a causal learning problem and propose Counterfactual Harness Intervention Learning for Long-Horizon Agents (CHILL-Harness). CHILL-Harness intervenes at the orchestration layer to enable advantage-guided workflow adaptation, thereby improving reasoning and execution efficiency while preserving task performance. Specifically, we develop causal intervention effect learning as the effect-estimation component of CHILL-Harness to estimate intervention-relative workflow advantage from confidence-weighted execution evidence and identify advantageous workflow adaptations. We further introduce advantage-realizing causal orchestration as its realization component to adaptively allocate counterfactual reasoning and realize only workflow adjustments supported by sufficient expected advantage. Finally, we incorporate a success-preserving objective and advantage-margin authorization constraints into CHILL-Harness to promote reliable adaptation. Extensive experiments on heterogeneous long-horizon tasks spanning information seeking, software engineering, and terminal interaction show that CHILL-Harness consistently preserves or improves task success while substantially reducing token consumption and execution time. Code: https://github.com/csdstar/chill-dev Introduction Large language model agents are increasingly deployed on long-horizon tasks Zhou et al. (2024); Xie et al. (2024) that require sustained context management Wang et al. (2025b), external tool interaction Yao et al. (2025), intermediate verification Pan et al. (2025), failure recovery Wang and Liu (2025), and execution control Guo et al. (2026). Agent harnesses have therefore become critical runtime infrastructure for translating latent model capability into reliable behavior Meng et al. (2026); Kapoor et al. (2026). However, effective harness control should adapt to task demands Hu et al. (2025d), execution environments Ba et al. (2026), and evolving execution states Zhang et al. (2026), whereas existing harnesses remain largely governed by hand-crafted rules, fixed thresholds, or globally configured workflows Li et al. (2026); Marchand et al. (2026). Over long execution horizons, this mismatch can compound redundant reasoning, delayed correction, and unnecessary workflow disruption, increasing execution cost and potentially degrading task success He et al. (2026); Sun et al. (2026). Existing harness-level approaches to adaptive harness orchestration can be broadly grouped into two categories: (1) task-oriented engineered harnesses such as Terminus-KIRA integrate hand-engineered reasoning and tool-use procedures across long-horizon tasks KRAFTON AI and Ludo Robotics (2026), OpenHands and CodeSweep–SWE-agent provide iterative repository inspection, editing, and verification workflows for software repair Wang et al. (2025a); Yang et al. (2024); Kimi Team (2025), and LemonHarness develops a high-performing execution stack for complex terminal tasks Ren et al. (2026); despite their strong task-level performance, these systems rely primarily on domain-specific or globally configured control logic and provide limited adaptation to the value of an intervention at the current execution state; and (2) orchestration and harness optimization methods improve system-level coordination—AWorld dynamically organizes multi-agent collaboration Yu et al. (2025a), OWL Workforce coordinates a planner, coordinator, and specialized workers through hierarchical orchestration Hu et al. (2025b), and Meta-Harness performs outer-loop search over harness implementations using execution traces and evaluation feedback Lee et al. (2026); however, these methods primarily optimize collaboration structures or aggregate harness configurations rather than determine whether a concrete orchestration intervention causally improves the ongoing workflow. Consequently, existing methods still lack a unified framework that assesses whether a concrete orchestration adjustment improves the current workflow and translates this assessment into adaptive, success-preserving harness decisions for efficient reasoning. In this work, we leverage a causal perspective to address this limitation. Causal reasoning is well suited to decisions whose observed outcomes conflate the effect of an action with the context in which it is taken Kuroki and Pearl (2014), since causal interventions can isolate the contribution of the action under a fixed context Rubin (1980); Pearl (2009). It has therefore been increasingly used to distinguish genuine contributions from spurious associations in large-language-model reasoning and autonomous driving Chi et al. (2024); Pourkeshavarz et al. (2024); Tang et al. (2026). Harness orchestration exhibits the same structure: the current workflow provides a factual reference, while admissible alternatives define counterfactual interventions under the same execution context. We therefore formulate the task of enabling adaptive orchestration in harness systems as a causal learning problem. Specifically, we introduce the causal harness orchestration problem: How can a harness leverage causal interventions to realize efficient and admissible workflow orchestration while preserving task success? To address this problem, we propose Counterfactual Harness Intervention Learning for Long-Horizon Agents (CHILL-Harness), a causal framework for adaptive harness orchestration and efficient long-horizon reasoning. CHILL-Harness intervenes at the orchestration layer to enable advantage-guided workflow adaptation. Specifically, we develop Causal Intervention Effect Learning (CIEL) as its effect-estimation component, which unifies heterogeneous orchestration behaviors and estimates their context-conditioned intervention-relative advantages from confidence-weighted execution evidence. We further introduce Advantage-Realizing Causal Orchestration (ARCO) as its realization component, which routes counterfactual deliberation before candidate generation and authorizes workflow adaptations only when they exhibit sufficient causal–operational advantage. Finally, we incorporate a success-preserving objective and advantage-margin authorization constraints to promote reliable orchestration under effect-estimation uncertainty. Extensive experiments on long-horizon tasks spanning information seeking, software engineering, and terminal interaction show that CHILL-Harness preserves or improves task success while substantially reducing token consumption and execution time. Our contributions are summarized as follows: • We propose CHILL-Harness, which formulates adaptive orchestration in harness systems as a causal learning problem over factual and admissible alternative workflows, enabling efficient reasoning through intervention. • We introduce Causal Intervention Effect Learning, which unifies heterogeneous orchestration behaviors and learns their context-conditioned intervention-relative advantages from confidence-weighted execution evidence, enabling selective harness-level intervention. • We develop Advantage-Realizing Causal Orchestration, which realizes workflow adaptations supported by sufficient estimated advantage, thereby achieving efficient and reliable harness orchestration under success-preserving constraints. Preliminaries and Related Work This section introduces the agent-harness setting and causal formulation, followed by a review of adaptive harness orchestration and causal intervention methods. Preliminaries: Causal Adaptation in Harnesses An agent harness is the runtime infrastructure that governs how a base language-model agent maintains context, invokes tools, verifies progress, and interacts with its task environment (Meng et al., 2026; Guo et al., 2026). At execution step t, the current context is represented as χt=(st,ctenv,gt), _t= (s_t,c_t^env,g_t ), where sts_t, ctenvc_t^env, and gtg_t denote the execution state, environment context, and current task objective, respectively. Given χt _t, the base agent proposes a factual workflow ωt0 _t^0, consisting of the reasoning, tool-use, and execution operations. Let tA_t denote the admissible workflow set under the current task, environment, safety, and resource constraints, with ωt0∈t _t^0 _t, and define tcf=t∖ωt0.A_t^cf=A_t \ _t^0\. The workflow selected for execution therefore satisfies Wt∈ωt0∪tcf.W_t∈\ _t^0\ _t^cf. Executing Wt=ωW_t=ω produces downstream task performance UtU_t under the common continuation and evaluation protocol. Adaptive harness orchestration determines whether to retain ωt0 _t^0 or execute an admissible alternative under χt _t. This decision naturally induces a potential-outcome formulation (Rubin, 1974; Pearl, 2009). For each ω∈tω _t, let Ut(ω)U_t(ω) denote the performance under do(Wt=ω)do(W_t=ω) with χt _t fixed. Under intervention consistency, Ut=Ut(Wt)U_t=U_t(W_t), and the workflow effect relative to the factual workflow is Γt(ω)=[Ut(ω)−Ut(ωt0)∣χt],Γt(ωt0)=0. _t(ω)=E [U_t(ω)-U_t( _t^0) _t ], _t( _t^0)=0. Related Work: Adaptive Harness Orchestration and Causal Intervention Existing work on adaptive harness orchestration includes engineered execution systems, automated harness optimization, and component-level adaptation. OpenHands, SWE-agent, and OWL improve execution through specialized interfaces or coordination (Wang et al., 2025a; Yang et al., 2024; Hu et al., 2025b); Automated Design of Agentic Systems, Multi-Agent Architecture Search, and Meta-Harness optimize agent structures, resources, or complete harnesses (Hu et al., 2025c; Zhang et al., 2025; Lee et al., 2026), while Agent Workflow Memory, HiAgent, and OSCAR adapt workflow reuse, context management, and recovery (Wang et al., 2025b; Hu et al., 2025a; Wang and Liu, 2025). Causal methods have also been explored for language-model reasoning and agent decision making. Causal Sufficiency and Necessity refines reasoning chains through counterfactual analysis (Yu et al., 2025b), Counterfactual Planning revises task-level actions using structural causal models (Fu et al., 2026), and Robust Agents Learn Causal World Models exploits causal environment structure for robust control (Richens and Everitt, 2024). These methods target reasoning steps, agent actions, or environment models, but do not assess heterogeneous harness-level decisions or efficiently realize only workflow-improving adaptations. CHILL-Harness fills this gap by estimating context-conditioned workflow effects and translating them into success-preserving workflow adaptations. CHILL-Harness: Counterfactual Harness Intervention Learning CHILL-Harness decomposes causal harness orchestration into two coupled problems. Definition 1 (Harness Intervention Effect Problem). The Harness Intervention Effect Problem evaluates whether an admissible workflow alternative improves execution relative to the factual workflow under the same execution context. Definition 2 (Harness Intervention Realization Problem). The Harness Intervention Realization Problem determines whether the estimated benefit of an admissible alternative is sufficient to justify replacing the factual workflow. Following the preliminaries, χt _t, ωt0 _t^0, tA_t, and Γt(ω) _t(ω) denote the execution context, factual workflow, admissible workflow set, and expected gain of ω over ωt0 _t^0, respectively. As illustrated in Figure 1, CHILL-Harness addresses the two problems through Causal Intervention Effect Learning (CIEL) and Advantage-Realizing Causal Orchestration (ARCO), under explicit success-preserving workflow authorization constraints. Figure 1: Overview of CHILL-Harness. Given context χt _t and factual workflow ωt0 _t^0, CIEL estimates workflow effects and predicts intervention, deliberation, revision, and completion signals. ARCO realizes them through route-before-generation, admissibility filtering, causal–operational valuation, and conservative authorization. Offline checkpointed paired executions supervise workflow-effect estimation, while trajectory outcomes provide weak supervision for the factorized decision heads. CIEL: Causal Intervention Effect Learning CIEL learns workflow-grounded intervention preferences for the Harness Intervention Effect Problem and converts them into factorized deliberation, revision, and completion signals. Unified Causal Intervention Induction. CIEL abstracts recurring harness adaptations, including additional reasoning, workflow correction, execution recovery, and answer construction, into a unified intervention space (Wang et al., 2025b; Yang et al., 2024; Wang and Liu, 2025). Let ℛ=Inspect,Verify,Visual,=Dedup,NoOpRecovery aligned R&=\ Inspect, Verify, Visual\,&Q&=\ Dedup, NoOpRecovery\ aligned denote the revision and stabilization modes. The intervention space and its workflow realization are ℐtwf= _t^wf= Keep,Deliberate,AnswerSynthesis \ Keep, Deliberate, AnswerSynthesis\ ∪Revise(r)∣r∈ℛ∪Stabilize(q)∣q∈, ∪\ Revise(r) r \∪\ Stabilize(q) q \, t(ι)≡ _t( )≡ (ι;χt,ωt0)⊆t,t(Keep)=ωt0. ( ; _t, _t^0 ) _t, _t( Keep)=\ _t^0\. Here, ℐtwfI_t^wf specifies adaptation intent and t(ι)G_t( ) maps it to admissible workflows. Keep preserves the factual workflow; Deliberate requests additional reasoning; Revise acquires, verifies, or visually inspects evidence; Stabilize suppresses repetition or recovers ineffective execution; and AnswerSynthesis redirects execution toward nonterminal answer construction. Each intervention inherits its causal contribution from its realized workflow (details in Appendix A.2). CIEL estimates the context-conditioned workflow effect as Γ^ϕ(ω;χt,ωt0)≈Γt(ω) _φ(ω; _t, _t^0)≈ _t(ω), with the factual workflow serving as the zero-effect reference. The estimates induce intervention-family preferences that are amortized into online prediction: V^tint(ι) V_t^int( ) =maxω∈t(ι)Γ^ϕ(ω;χt,ωt0), = _ω _t( ) _φ (ω; _t, _t^0 ), (1a) ιt∗ _t^* =argmaxι∈ℐtwfV^tint(ι), = _ _t^wf V_t^int( ), (1b) ι^t _t =argmaxι∈ℐtwfpθu(ι∣χt,ωt0). = _ _t^wfp_ _u ( _t, _t^0 ). (1c) Equations (1a)–(1c) separate workflow-effect aggregation, offline intervention induction, and amortized online prediction. Because t(Keep)=ωt0G_t( Keep)=\ _t^0\, conservative tie-breaking selects Keep whenever no alternative family has a positive estimated effect. The workflow-effect estimator learns from checkpointed paired records, while the intervention predictor amortizes the induced offline preference: ℒeff _eff =pair[wt,ω(Γ^ϕ(ω;χt,ωt0)−Γ~tpair(ω))2], =E_D_pair [w_t,ω ( _φ (ω; _t, _t^0 )- _t^pair(ω) )^2 ], ℒint _int =−logpθu(ιt∗∣χt,ωt0). =- p_ _u ( _t^* _t, _t^0 ). Here, pairD_pair contains factual–candidate executions restored from the same context, Γ~tpair(ω) _t^pair(ω) is their empirical utility difference, and wt,ωw_t,ω is its reliability weight. For analysis, define Vtint(ι)=maxω∈t(ι)Γt(ω),V_t^int( )= _ω _t( ) _t(ω), ιt†=argmaxι∈ℐtwfVtint(ι), _t = _ _t^wfV_t^int( ), and let t=⋃ι∈ℐtwft(ι).C_t= _ _t^wfG_t( ). The workflow-effect error and amortization gap are εt=supω∈t|Γ^ϕ(ω;χt,ωt0)−Γt(ω)|,ρt=V^tint(ιt∗)−V^tint(ι^t). aligned _t&= _ω _t | _φ (ω; _t, _t^0 )- _t(ω) |,& _t&= V_t^int( _t^*)- V_t^int( _t). aligned (2) Then Vtint(ιt†)−Vtint(ι^t)≤2εt+ρt.V_t^int( _t )-V_t^int( _t)≤ 2 _t+ _t. (3) Thus, effect-estimation and amortization errors jointly control intervention selection regret; Appendix A.3 proves the result, and Appendix A.4 details paired-effect construction. Factorized Intervention Decision Learning. CIEL factorizes online control into three decisions governing additional deliberation, workflow revision, and task completion. Counterfactual Deliberation Effect (CDE). CDE addresses harness over-deliberation, where costly planning is repeatedly invoked even when the factual workflow is already sufficient. It predicts the required deliberation route: pθp(zt∣χt,ωt0,ι^t),zt∈=Skip,Light,Full.p_ _p (z_t _t, _t^0, _t ), z_t =\ Skip, Light, Full\. (4) Skip bypasses candidate generation, Light enables restricted diagnostics, and Full enables broader evaluation. Thus, the intervention family specifies what adaptation is useful, whereas CDE determines how much computation realizes it. Counterfactual Revision Effect (CRE). CRE addresses both workflow inertia and harmful replacement: the factual workflow may require correction, while an unsupported alternative may further degrade execution. Given an alternative workflow ω, CRE predicts whether it provides sufficient support for replacing the factual workflow: pθc(ytchg∣χt,ωt0,ω,ι^t),ytchg∈Keep,Change.p_ _c (y_t^chg _t, _t^0,ω, _t ), y_t^chg∈\ Keep, Change\. (5) Keep retains ωt0 _t^0, whereas Change supports replacement by the evaluated alternative. Completion Attribution Effect (CAE). CAE predicts completion support to prevent both continued execution after task success and termination triggered solely by futility: pθe(bt∣χt,ot,ht),bt∈0,1.p_ _e (b_t _t,o_t,h_t ), b_t∈\0,1\. (6) Here, oto_t and hth_t denote the latest observation and recent execution summary; bt=1b_t=1 supports completion, which still requires explicit task evidence. Logged execution traces provide weak targets zt∗z_t^*, ytchg,∗y_t^chg,*, and bt∗b_t^* for CDE, CRE, and CAE, respectively. Here, zt∗z_t^* denotes the preferred deliberation route, ytchg,∗y_t^chg,* denotes the preferred revision decision, and bt∗b_t^* denotes whether retrospective task evidence supports completion. Efficiency-related decisions receive positive supervision only when task success is preserved. The corresponding losses are ℒCDE _CDE =−logpθp(zt∗∣χt,ωt0,ι^t), =- p_ _p (z_t^* _t, _t^0, _t ), ℒCRE _CRE =−logpθc(ytchg,∗∣χt,ωt0,ωt+,ι^t), =- p_ _c (y_t^chg,* _t, _t^0, _t^+, _t ), ℒCAE _CAE =−logpθe(bt∗∣χt,ot,ht). =- p_ _e (b_t^* _t,o_t,h_t ). Thus, ℒCDEL_CDE supervises the amount of additional reasoning, ℒCREL_CRE supervises factual-versus-candidate workflow selection, and ℒCAEL_CAE supervises evidence-attributed completion. The complete CIEL objective is ℒCIEL=ℒeff+ℒint+ℒCDE+ℒCRE+ℒCAE.L_CIEL=L_eff+L_int+L_CDE+L_CRE+L_CAE. (7) Equation (7) jointly supervises workflow effects, intervention preferences, deliberation, revision, and completion. Detailed weak-target construction is provided in Appendix A.4. Remark. CIEL separates what adaptation is beneficial from how it should be realized: workflow-effect learning induces intervention preferences, while CDE, CRE, and CAE factorize deliberation, revision, and completion decisions. ARCO: Advantage-Realizing Causal Orchestration ARCO addresses the Harness Intervention Realization Problem through route selection, candidate valuation, conservative revision, and evidence-grounded completion. Let Γ^ϕ,t(ω):=Γ^ϕ(ω;χt,ωt0) _φ,t(ω):= _φ(ω; _t, _t^0). Then ARCO(χt,ωt0,ot,ht,ι^t,Γ^ϕ,t(⋅))⟼(ωt⋆,dtterm).ARCO ( _t, _t^0,o_t,h_t, _t, _φ,t(·) ) ( _t ,d_t^term ). Using the CDE predictor defined in Equation (4), ARCO first selects the provisional deliberation route: z~t=argmaxz∈pθp(z∣χt,ωt0,ι^t). z_t= _z p_ _p (z _t, _t^0, _t ). Under Skip≺Light≺Full, Skip Light Full, ztz_t is the highest feasible route not exceeding z~t z_t. Skip retains ωt0 _t^0; otherwise, the selected route exposes Ωt⊆ωt0∪t(ι^t),Ω¯t=Ωt∩t, _t \ _t^0\ _t( _t), _t= _t _t, where Ω¯t _t contains the admissible route-exposed workflows. Route feasibility is detailed in Appendix B.1. Each admissible candidate is valued by combining operational utility and estimated workflow effect: U^ARCO(ω∣χt,zt,Γ^ϕ,t)=U^op(ω∣χt,zt)+ηΓΓ^ϕ,t(ω). U_ARCO (ω _t,z_t, _φ,t )= U_op (ω _t,z_t )+ _ _φ,t(ω). (8) Here, U^op U_op aggregates progress, cost, risk, information gain, robustness, and safety, while ηΓ≥0 _ ≥ 0 weights estimated intervention evidence; Appendix B.2 gives the full construction. ARCO selects the highest-valued workflow among the admissible route-exposed candidates: ωt+=argmaxω∈Ω¯tU^ARCO(ω∣χt,zt,Γ^ϕ,t), _t^+= _ω∈ _t U_ARCO (ω _t,z_t, _φ,t ), with advantage over the factual workflow Δtcf=U^ARCO(ωt+∣χt,zt,Γ^ϕ,t)−U^ARCO(ωt0∣χt,zt,Γ^ϕ,t). aligned _t^cf= U_ARCO ( _t^+ _t,z_t, _φ,t )- U_ARCO ( _t^0 _t,z_t, _φ,t ). aligned Instantiating the CRE predictor in Equation (5) with the selected candidate ωt+ _t^+, ARCO obtains ytchg=argmaxy∈Keep,Changepθc(y∣χt,ωt0,ωt+,ι^t).y_t^chg= _y∈\ Keep, Change\p_ _c (y _t, _t^0, _t^+, _t ). The candidate is executed only when both CRE and its estimated advantage support revision: ωt⋆=ωt+,ytchg=Change∧Δtcf≥δι^t,ωt0,otherwise. _t = cases _t^+,&y_t^chg= Change\ \ _t^cf≥ _ _t,\\[2.84526pt] _t^0,&otherwise. cases (9) In Equation (9), δι^t≥0 _ _t≥ 0 is an intervention-dependent authorization margin. ARCO executes ωt+ _t^+ only when both the CRE decision and its estimated advantage support replacement; otherwise, it retains ωt0 _t^0 (see Appendix B.3 for details). Finally, ARCO instantiates the CAE predictor in Equation (6) and combines its output with explicit task evidence: b^t b_t =argmaxb∈0,1pθe(b∣χt,ot,ht), = _b∈\0,1\p_ _e (b _t,o_t,h_t ), etcomp e_t^comp =[t(χt,ot,ht)=1], =I [V_t ( _t,o_t,h_t )=1 ], dtterm d_t^term =b^tetcomp. = b_te_t^comp. Here, tV_t is the task verifier and [⋅]I[·] is the binary indicator. Remark. ARCO realizes learned effects conservatively: routing limits candidate-generation cost, causal–operational valuation ranks admissible alternatives, and advantage-margin authorization retains the factual workflow when evidence is insufficient. Success-Preserving Objective and Authorization Constraints Let τ be a complete execution trajectory, πCHILL,Θ _CHILL, the policy induced by CIEL and ARCO, Θ=ϕ,θu,θp,θc,θe =\φ, _u, _p, _c, _e\, and πref _ref the matched reference harness. CHILL-Harness minimizes trajectory cost subject to bounded success degradation: minΘ _ τ∼πCHILL,Θ[eff(τ)] _τ _CHILL, [J_eff(τ) ] s.t. PassRate(πCHILL,Θ)≥PassRate(πref)−ϵpass, ( _CHILL, ) ( _ref )- _pass, where eff(τ)J_eff(τ) aggregates trajectory-level resource cost and ϵpass≥0 _pass≥ 0 is the allowed pass-rate degradation. At each step, only admissible route-exposed workflows may be executed, and termination requires both learned and verified completion evidence: ωt⋆∈Ω¯t=Ωt∩t,dtterm=b^tetcomp. _t ∈ _t= _t _t, d_t^term= b_te_t^comp. Because environment execution is not directly differentiable, the constrained objective is realized through outcome-conditioned weak supervision rather than trajectory-level backpropagation. Let Yτ,Yref∈0,1Y_τ,Y_ref∈\0,1\ denote the current and reference trajectory outcomes, and let λfail,λbreak≥0 _fail, _break≥ 0 be their penalty weights. We define ℒtraj(τ)=eff(τ)+λfail(1−Yτ)+λbreak[Yref=1∧Yτ=0]. aligned L_traj(τ)=J_eff(τ)+ _fail (1-Y_τ )+ _breakI [Y_ref=1 Y_τ=0 ]. aligned (10) Equation (10) penalizes resource cost, task failure, and success-breaking adaptation, and constructs weak targets for CDE, CRE, and CAE. Remark. The success-preserving objective and authorization constraints jointly protect task performance: efficiency-related decisions receive positive supervision only when success is preserved, while unsupported workflow revision and premature termination are disallowed. Algorithm 1 summarizes the inference flow; Appendices C.1–C.4 provide the complete objectives and authorizations. Algorithm 1 CHILL-Harness Overview (Full Procedure in Appendix C.4) 1: Input: χt,ωt0,ot,ht _t, _t^0,o_t,h_t; Output: ωt⋆,dtterm _t ,d_t^term 2: (ι^t,z~t)←CIEL(χt,ωt0)( _t, z_t) ( _t, _t^0) 3: Set ztz_t to the highest feasible route not exceeding z~t z_t 4: if zt=Skipz_t= Skip then 5: Ωt←ωt0 _t←\ _t^0\ 6: else 7: Generate Ωt _t from t(ι^t)∪ωt0G_t( _t)∪\ _t^0\ 8: end if 9: Ω¯t←Ωt∩t _t← _t _t; select the highest-valued candidate ωt+ _t^+ 10: Authorize ωt⋆ _t using CRE and candidate advantage 11: Authorize dttermd_t^term using CAE and completion evidence 12: if dtterm=1d_t^term=1 then 13: Invoke the task-terminal action 14: else 15: Execute ωt⋆ _t 16: end if 17: Log execution evidence Benchmark Method Success Tokens Token Gain vs. Min.-Token Tokens/Solved Rel. Runtime Time Gain vs. Min.-Runtime GAIA Terminus-KIRA 70.2% 134M – 2.03M 1.00×1.00× – AWorld 69.7% 161M – 2.45M 1.25×1.25× – OWL Workforce 69.1% 174M – 2.68M 1.30×1.30× – CHILL-Harness 71.3% 96M 28.4%↓ 1.43M 0.534×0.534× 46.6%↓ SWE-bench Verified Terminus-KIRA 65.0% 157M – 2.42M 1.00×1.00× – OpenHands 65.4% 165M – 2.52M 1.10×1.10× – CodeSweep–SWE-agent 53.4% 153M – 2.87M 0.92×0.92× – CHILL-Harness 65.6% 133M 13.1%↓ 2.25M 0.682×0.682× 25.9%↓ Terminal- Bench 2.0 Terminus-KIRA 49.4% 291M – 7.11M 1.00×1.00× – Meta-Harness 52.6% 340M – 5.00M 1.25×1.25× – LemonHarness 57.4% 380M – 5.05M 1.40×1.40× – CHILL-Harness 56.1% 224M 23.1%↓ 4.87M 0.974×0.974× 2.6%↓ Table 1: Effectiveness and efficiency across three long-horizon benchmarks. Token Gain and Time Gain use the eligible non-CHILL baseline with the lowest token consumption and runtime, respectively. Relative runtime is normalized to Terminus-KIRA within each benchmark. Token totals are rounded to the nearest million, whereas Tokens/Solved is reported to two decimal places. CHILL-Harness token totals include all additional model calls for counterfactual deliberation and candidate valuation. Experiments We evaluate whether CHILL-Harness resolves the causal harness orchestration problem by learning beneficial workflow interventions, realizing them efficiently and admissibly, and preserving task success. We examine these requirements through three empirical questions: • RQ1: Effectiveness. Does solving the Harness Intervention Effect Problem in Definition 1 identify adaptations that preserve or improve task success? • RQ2: Efficiency. Does solving the Harness Intervention Realization Problem in Definition 2 reduce reasoning and execution cost without premature failure or termination? • RQ3: Generalization. Do the effectiveness and efficiency benefits hold across heterogeneous execution environments? We conduct baseline comparisons to verify that CHILL-Harness preserves effectiveness while improving efficiency across benchmarks, and perform an Always-Full ablation to establish the contribution of counterfactual orchestration. Experimental Setup Benchmarks. We evaluate CHILL-Harness on GAIA for information seeking and tool-assisted reasoning (Mialon et al., 2024), SWE-bench Verified for repository-level software repair (Jimenez et al., 2024; OpenAI, 2025), and Terminal-Bench 2.0 for long-horizon terminal interaction (Merrill et al., 2026). These benchmarks cover deliberation, revision, stabilization, verification, and completion under distinct execution environments. Baselines. Each CHILL-Harness evaluation is paired with a reference run using the same model, tools, environment, task, and resource limits. For external comparison, we report Terminus-KIRA on all benchmarks (KRAFTON AI and Ludo Robotics, 2026); AWorld and OWL Workforce on GAIA (Yu et al., 2025a; Hu et al., 2025b); OpenHands and CodeSweep–SWE-agent on SWE-bench Verified (Wang et al., 2025a; Yang et al., 2024; Kimi Team, 2025); and Meta-Harness and LemonHarness on Terminal-Bench 2.0 (Lee et al., 2026; Ren et al., 2026). Figure 2: Joint effectiveness–efficiency comparison on GAIA, SWE-bench Verified, and Terminal-Bench 2.0. The horizontal axis reports task success and the vertical axis reports solved tasks per one million tokens; higher values on both axes are preferred. Stars denote CHILL-Harness, circles denote public baseline harnesses, and colors group methods evaluated on the same benchmark. Model Configuration. GAIA uses deepseek-v4-flash, whereas SWE-bench Verified and Terminal-Bench 2.0 use deepseek-v4-pro DeepSeek-AI (2026). CIEL is trained offline on disjoint paired traces and frozen for evaluation; checkpointed replay is disabled at test time. All model calls use temperature 1.01.0 and at most 150150 interaction turns. Evaluation Metrics. Let m denote a method, N the number of tasks, and yi(m)∈0,1y_i^(m)∈\0,1\ whether m solves task i. Effectiveness is measured by Success(m)=1N∑i=1Nyi(m)×100%.Success(m)= 1N _i=1^Ny_i^(m)× 100\%. Reasoning cost is measured by Tokens(m)=∑i=1N∑k∈m(Toki,kin+Toki,kout),Tokens/Solved(m)=Tokens(m)∑i=1Nyi(m),Tokens(m)= _i=1^N _k _m (Tok^in_i,k+Tok^out_i,k ),Tokens/Solved(m)= Tokens(m) _i=1^Ny_i^(m), where mK_m includes all model-based harness components. Let Ti(m)T_i^(m) be the end-to-end runtime of method m on task i, including inference, orchestration, tools, interaction, and verification: RelativeRuntime(m;bT)=∑i=1NTi(m)∑i=1NTi(bT).RelativeRuntime(m;b_T)= _i=1^NT_i^(m) _i=1^NT_i^(b_T). For X∈Tokens,TX∈\Tokens,T\, resource gain is GainX(m;bX)=(1−X(m)X(bX))×100%.Gain_X(m;b_X)= (1- X(m)X(b_X) )× 100\%. Here, bTb_T and bXb_X denote the runtime and resource reference baselines, respectively. Benchmark Variant Success Tokens Tokens/Solved Full Calls / Revision Rate Always-Full Overhead GAIA CHILL-Harness 71.3% 96.00M 1.43M – – Always-Full 27.7% 114.33M 4.40M 2,615 / 76.1% 18.33M tokens 37.73 h SWE-bench Verified CHILL-Harness 65.6% 133.00M 2.25M – – Always-Full 53.3% 160.40M 3.34M 3,708 / 3.5% 27.40M tokens 70.95 h Terminal- Bench 2.0 CHILL-Harness 56.1% 224.19M 4.87M – – Always-Full 56.3% 259.77M 5.65M 4,234 / 3.9% 35.58M tokens 91.98 h Table 2: Ablation Experiment Results. Full Calls / Revision Rate reports the number of full CDE events and the fraction of those events that authorize workflow revision. Always-Full Overhead reports the additional planner token consumption and cumulative planner latency incurred by Always-Full relative to CHILL-Harness on the same benchmark. Main Results Table 1 reports effectiveness and resource measurements against baselines, while Figure 2 summarizes their joint success–token efficiency. Controlled matched-reference comparisons assess success preservation, whereas comparisons with public systems assess external competitiveness. RQ1: Effectiveness. Under matched model, tool, environment, task, and resource settings, CHILL-Harness preserves or improves task success on all three benchmarks. It achieves the highest success rate among the compared systems on GAIA (71.3%71.3\%) and SWE-bench Verified (65.6%65.6\%). On Terminal-Bench 2.0, its 56.1%56.1\% success exceeds Terminus-KIRA and Meta-Harness and remains close to LemonHarness (57.4%57.4\%). Thus, the efficiency gains do not arise from broadly sacrificing task effectiveness. RQ2: Efficiency. CHILL-Harness achieves the highest number of solved tasks per one million tokens in each benchmark group. Against the lowest-token public baselines, it reduces token consumption by 28.4%28.4\%, 13.1%13.1\%, and 23.1%23.1\% on GAIA, SWE-bench Verified, and Terminal-Bench 2.0, respectively. Against the fastest eligible baselines, it reduces runtime by 46.6%46.6\%, 25.9%25.9\%, and 2.6%2.6\%. RQ3: Generalization. CHILL-Harness improves both success and token efficiency over all displayed baselines on GAIA and SWE-bench Verified. On Terminal-Bench 2.0, it achieves the highest token efficiency, while LemonHarness retains a small success advantage, placing both systems on the effectiveness–efficiency frontier. This pattern holds across information seeking, software repair, and terminal interaction. Remark. Controlled comparisons support success preservation, while public-system comparisons show that CHILL-Harness consistently improves token efficiency and remains competitive in task effectiveness across heterogeneous long-horizon environments. Ablation Experiment We evaluate the contribution of CDE and its route-before-generation realization in ARCO. The Always-Full variant disables adaptive deliberation by setting zt=Fullz_t= Full whenever feasible, while retaining CIEL intervention prediction, CRE, and CAE, as well as ARCO candidate generation, admissibility filtering, valuation, and authorization. The comparison therefore isolates the effect of selecting deliberation depth before candidate generation. Table 2 shows that forced full deliberation causes over-intervention on GAIA, increases cost while reducing success on SWE-bench Verified, and incurs substantial planner overhead with little benefit on Terminal-Bench 2.0. These results validate the contribution of CDE and its route-before-generation realization in ARCO. Remark. CDE acts before candidate generation: CRE can reject unsupported workflow revisions only after the corresponding candidate-generation cost has been incurred, whereas route-before-generation suppresses low-value deliberation and avoids the associated planner overhead. Conclusion We introduced CHILL-Harness, a causal framework that learns whether and how the harness should intervene in the choice between factual and alternative workflows. The empirical results support affirmative answers to our three research questions: CIEL identifies interventions that preserve or improve task effectiveness, ARCO realizes them with lower reasoning and execution cost, and these benefits generalize across heterogeneous long-horizon environments. Experiments on information seeking, software repair, and terminal interaction, together with the Always-Full ablation, support both the overall effectiveness of CHILL-Harness and the necessity of selective counterfactual deliberation. We hope this work encourages the community to move beyond fixed harness engineering toward learned, causal, and success-preserving orchestration, providing a foundation for more efficient and reliable long-horizon agents. References L. Ba, Q. Li, and S. Li (2026) Ciber: a comprehensive benchmark for security evaluation of code interpreter agents. arXiv:2602.19547. Cited by: Introduction. H. Chi, H. Li, W. Yang, F. Liu, L. Lan, X. Ren, T. Liu, and B. Han (2024) Unveiling causal reasoning in large language models: reality or mirage?. NeurIPS. Cited by: Introduction. DeepSeek-AI (2026) DeepSeek-v4: towards highly efficient million-token context intelligence. Cited by: Model Configuration.. J. Fu, L. Ding, Q. Wei, Y. Guo, Y. Cheng, and J. Zhang (2026) Counterfactual planning for generalizable agents’ actions. AAAI. Cited by: Related Work: Adaptive Harness Orchestration and Causal Intervention. J. Guo, Z. Hao, C. Wang, C. Fan, T. Luo, H. Li, Y. Gao, H. Mei, J. Peng, R. Xu, et al. (2026) From question answering to task completion: a survey on agent system and harness design. arXiv:2606.20683. Cited by: Introduction, Preliminaries: Causal Adaptation in Harnesses. Z. He, Y. Wang, C. Zhi, Y. Hu, T. Chen, L. Yin, Z. Chen, T. A. Wu, S. Ouyang, Z. Wang, et al. (2026) MemoryArena: benchmarking agent memory in interdependent multi-session agentic tasks. arXiv:2602.16313. Cited by: Introduction. M. Hu, T. Chen, Q. Chen, Y. Mu, W. Shao, and P. Luo (2025a) Hiagent: hierarchical working memory management for solving long-horizon agent tasks with large language model. ACL. Cited by: Related Work: Adaptive Harness Orchestration and Causal Intervention. M. Hu, Y. Zhou, W. Fan, Y. Nie, B. Xia, T. Sun, Z. Ye, Z. Jin, Y. Li, Q. Chen, Z. Zhang, Y. Wang, Q. Ye, B. Ghanem, P. Luo, and G. Li (2025b) OWL: optimized workforce learning for general multi-agent assistance in real-world task automation. NeurIPS. Cited by: Introduction, Related Work: Adaptive Harness Orchestration and Causal Intervention, Baselines.. S. Hu, C. Lu, and J. Clune (2025c) Automated design of agentic systems. ICLR. Cited by: Related Work: Adaptive Harness Orchestration and Causal Intervention. Y. Hu, Y. Wang, and J. McAuley (2025d) Evaluating memory in llm agents via incremental multi-turn interactions. arXiv:2507.05257. Cited by: Introduction. C. E. Jimenez, J. Yang, A. Wettig, S. Yao, K. Pei, O. Press, and K. R. Narasimhan (2024) SWE-bench: can language models resolve real-world github issues?. ICLR. Cited by: Benchmarks.. S. Kapoor, B. Stroebl, P. Kirgis, N. Nadgir, Z. S. Siegel, B. Wei, T. Xue, Z. Chen, F. Chen, S. Utpala, F. Ndzomga, D. Oruganty, S. Luskin, K. Liu, B. Yu, A. Arora, D. Hahm, H. Trivedi, H. Sun, J. Lee, T. Jin, Y. Mai, Y. Zhou, Y. Zhu, R. Bommasani, D. Kang, D. Song, P. Henderson, Y. Su, P. Liang, and A. Narayanan (2026) Holistic agent leaderboard: the missing infrastructure for ai agent evaluation. ICLR. Cited by: Introduction. Kimi Team (2025) Kimi K2: open agentic intelligence. arXiv:2507.20534. Cited by: Introduction, Baselines.. KRAFTON AI and Ludo Robotics (2026) Terminus-kira: boosting frontier model performance on terminal-bench with minimal harness. External Links: Link Cited by: Introduction, Baselines.. M. Kuroki and J. Pearl (2014) Measurement bias and effect restoration in causal inference. Biometrika 101, p. 423–437. Cited by: Introduction. Y. Lee, R. Nair, Q. Zhang, K. Lee, O. Khattab, and C. Finn (2026) Meta-harness: end-to-end optimization of model harnesses. arXiv:2603.28052. Cited by: Introduction, Related Work: Adaptive Harness Orchestration and Causal Intervention, Baselines.. J. Li, X. Xiao, Y. Zhang, C. Liu, L. Zhao, X. Liao, Y. Ji, J. Wang, Y. Ge, W. Xu, X. Fang, X. Xu, T. Zhao, Y. Kim, J. Hamm, T. Wang, and C. Reddy (2026) Agent harness engineering: a survey. External Links: Link Cited by: Introduction. R. Marchand, A. O. Cathain, J. Wynne, P. M. Giavridis, S. Deverett, J. Wilkinson, J. Gwartz, and H. Coppock (2026) Quantifying frontier llm capabilities for container sandbox escape. ICML. Cited by: Introduction. Q. Meng, Y. Wang, L. Chen, W. Wu, Y. Li, W. Jiang, Q. Wang, C. Lu, Y. Gao, Y. Wu, and Y. Hu (2026) Agent harness for large language model agents: a survey. Preprints:0.20944/preprints202604.0428.v3. Cited by: Introduction, Preliminaries: Causal Adaptation in Harnesses. M. A. Merrill, A. G. Shaw, and e. Nicholas Carlini (2026) Terminal-bench: benchmarking agents on hard, realistic tasks in command line interfaces. ICLR. Cited by: Benchmarks.. G. Mialon, C. Fourrier, C. Swift, T. Wolf, Y. LeCun, and T. Scialom (2024) GAIA: a benchmark for general ai assistants. ICLR. Cited by: Benchmarks.. OpenAI (2025) Introducing SWE-bench Verified. Note: OpenAI Research External Links: Link Cited by: Benchmarks.. J. Pan, X. Wang, G. Neubig, N. Jaitly, H. Ji, A. Suhr, and Y. Zhang (2025) Training software engineering agents and verifiers with swe-gym. ICML. Cited by: Introduction. J. Pearl (2009) Causality. Cambridge university press. Cited by: Introduction, Preliminaries: Causal Adaptation in Harnesses. M. Pourkeshavarz, J. Zhang, and A. Rasouli (2024) CaDeT: a causal disentanglement approach for robust trajectory prediction in autonomous driving. CVPR. Cited by: Introduction. K. Ren, F. Sun, J. Liu, L. Yang, Z. Yin, J. Li, C. Yin, M. He, Y. Huo, J. Liu, Z. Chen, Y. Huangfu, R. Li, Y. Wu, X. Su, Y. Xu, L. Wu, H. Zhao, L. Zhang, X. Geng, and J. Fan (2026) LemonHarness technical report. arXiv:2606.24311. Cited by: Introduction, Baselines.. J. Richens and T. Everitt (2024) Robust agents learn causal world models. ICLR. Cited by: Related Work: Adaptive Harness Orchestration and Causal Intervention. D. B. Rubin (1974) Estimating causal effects of treatments in randomized and nonrandomized studies. Journal of Educational Psychology 66, p. 688–701. Cited by: Preliminaries: Causal Adaptation in Harnesses. D. B. Rubin (1980) Randomization analysis of experimental data: the fisher randomization test comment. Journal of the American Statistical Association 75, p. 591. Cited by: Introduction. S. Sun, H. Song, L. Huang, J. Jiang, R. Le, Z. Lv, Z. Chen, Y. Hu, W. Luo, W. X. Zhao, Y. Song, H. Xu, T. Zhang, and J. Wen (2026) SWE-world: building software engineering agents in docker-free environments. arXiv:2602.03419. Cited by: Introduction. J. Tang, Z. Zhou, Z. He, J. Zhang, K. Zhang, and J. Pu (2026) CausalVAD: de-confounding end-to-end autonomous driving via causal intervention. CVPR. Cited by: Introduction. X. Wang and B. Liu (2025) Oscar: operating system control via state-aware reasoning and re-planning. Cited by: Introduction, Related Work: Adaptive Harness Orchestration and Causal Intervention, Unified Causal Intervention Induction.. X. Wang, B. Li, Y. Song, F. F. Xu, X. Tang, M. Zhuge, J. Pan, Y. Song, B. Li, J. Singh, H. H. Tran, F. Li, R. Ma, M. Zheng, B. Qian, Y. Shao, N. Muennighoff, Y. Zhang, B. Hui, J. Lin, R. Brennan, H. Peng, H. Ji, and G. Neubig (2025a) OpenHands: an open platform for ai software developers as generalist agents. ICLR. Cited by: Introduction, Related Work: Adaptive Harness Orchestration and Causal Intervention, Baselines.. Z. Z. Wang, J. Mao, D. Fried, and G. Neubig (2025b) Agent workflow memory. ICML. Cited by: Introduction, Related Work: Adaptive Harness Orchestration and Causal Intervention, Unified Causal Intervention Induction.. T. Xie, D. Zhang, J. Chen, X. Li, S. Zhao, R. Cao, T. J. Hua, Z. Cheng, D. Shin, F. Lei, et al. (2024) Osworld: benchmarking multimodal agents for open-ended tasks in real computer environments. NeurIPS. Cited by: Introduction. J. Yang, C. E. Jimenez, A. Wettig, K. Lieret, S. Yao, K. R. Narasimhan, and O. Press (2024) SWE-agent: agent-computer interfaces enable automated software engineering. NeurIPS. Cited by: Introduction, Related Work: Adaptive Harness Orchestration and Causal Intervention, Unified Causal Intervention Induction., Baselines.. S. Yao, N. Shinn, P. Razavi, and K. Narasimhan (2025) au tau-Bench: a benchmark for tool-agent-user interaction in real-world domains. ICLR. Cited by: Introduction. C. Yu, S. Lu, and C. Zhuang,etc (2025a) AWorld: orchestrating the training recipe for agentic ai. arXiv:2508.20404. Cited by: Introduction, Baselines.. X. Yu, Z. Wang, L. Yang, H. Li, A. Liu, X. Xue, J. Wang, and M. Yang (2025b) Causal sufficiency and necessity improves chain-of-thought reasoning. NeurIPS. Cited by: Related Work: Adaptive Harness Orchestration and Causal Intervention. G. Zhang, L. Niu, J. Fang, K. Wang, L. Bai, and X. Wang (2025) Multi-agent architecture search via agentic supernet. ICML. Cited by: Related Work: Adaptive Harness Orchestration and Causal Intervention. Y. Zhang, J. Wang, Y. Ge, W. Xu, J. Hamm, and C. K. Reddy (2026) Stop comparing llm agents without disclosing the harness. arXiv:2605.23950. Cited by: Introduction. S. Zhou, F. F. Xu, H. Zhu, X. Zhou, R. Lo, A. Sridhar, X. Cheng, T. Ou, Y. Bisk, D. Fried, et al. (2024) Webarena: a realistic web environment for building autonomous agents. ICLR. Cited by: Introduction. Appendix A: Problem Formalization and CIEL Foundations This appendix provides the formal definitions, theoretical grounding, and training implementation of Causal Intervention Effect Learning (CIEL). Appendix A.1 formalizes the two harness intervention problems. Appendix A.2 specifies the structural connection between semantic interventions and executable workflows. Appendix A.3 derives an error-dependent bound for amortized intervention selection. Appendix A.4 establishes paired-replay identification and details the separated supervision of workflow effects and factorized decisions. A.1 Formalization of the Harness Intervention Problems At execution step t, let χt=(st,ctenv,gt) _t= (s_t,c_t^env,g_t ) denote the fixed execution context, where sts_t is the execution state, ctenvc_t^env is the environment context, and gtg_t is the current task objective. Let ωt0 _t^0 denote the factual workflow proposed without additional harness adaptation, and let tA_t denote the workflows admissible under the current task, environment, safety, and resource requirements, with ωt0∈t. _t^0 _t. The executed workflow is represented by WtW_t, and its task- and resource-aware execution performance is represented by UtU_t. For any ω∈tω _t, do(Wt=ω)do(W_t=ω) denotes executing workflow ω while holding χt _t fixed. All potential outcomes are evaluated under the same continuation policy πcontπ^cont, evaluation horizon, resource-accounting rule, and task evaluator. Ut(ω)U_t(ω) captures downstream task performance, whereas local cost, risk, and execution-quality signals are represented separately by U^op U_op. Definition A.1: Harness Intervention Effect Problem. For an admissible workflow ω∈t∖ωt0,ω _t \ _t^0\, its intervention effect relative to the factual workflow is Γt(ω)=[Ut∣do(Wt=ω),χt]−[Ut∣do(Wt=ωt0),χt]. aligned _t(ω)=E [U_t (W_t=ω), _t ]-E [U_t (W_t= _t^0), _t ]. aligned (A1) A positive value indicates expected improvement relative to the factual workflow, whereas a negative value indicates expected degradation. The factual workflow is the zero-effect reference: Γt(ωt0)=0. _t( _t^0)=0. The estimand therefore assigns causal value to concrete executable workflows rather than directly to abstract intervention labels. Definition A.2: Harness Intervention Realization Problem. Let Γ^ϕ,t(ω):=Γ^ϕ(ω;χt,ωt0) _φ,t(ω):= _φ (ω; _t, _t^0 ) denote the estimated workflow effect, and let Ω¯t⊆t _t _t be the route-exposed admissible workflow set with ωt0∈Ω¯t _t^0∈ _t. The highest-valued exposed candidate is ωt+=argmaxω∈Ω¯tU^ARCO(ω∣χt,zt,Γ^ϕ,t). _t^+= _ω∈ _t U_ARCO (ω _t,z_t, _φ,t ). (A2) The authorized workflow then satisfies ωt⋆∈ωt0,ωt+. _t ∈ \ _t^0, _t^+ \. (A3) Candidate selection and execution authorization are distinct: selection identifies the strongest currently exposed candidate, whereas replacement additionally requires admissibility, learned revision support, and a sufficient estimated advantage. The resulting orchestration chain is workflow-effect estimation⟶route-exposed valuation⟶execution authorization.workflow-effect estimation -exposed valuation authorization. A.2 Workflow-Grounded Intervention Induction Let It∈ℐtwfI_t _t^wf denote the semantic harness intervention selected at step t. The intervention families have the following meanings. • Keep preserves the factual workflow ωt0 _t^0. • Deliberate allocates additional reasoning without prescribing a particular workflow revision. • Revise(Inspect) acquires execution evidence missing from the current context. • Revise(Verify) validates an intermediate state, tool result, or candidate answer. • Revise(Visual) invokes visual processing when image or multimodal evidence is required. • Stabilize(Dedup) suppresses repeated or semantically redundant operations. • Stabilize(NoOpRecovery) redirects an execution that has produced an empty, ineffective, or stalled action. • AnswerSynthesis redirects broad exploration toward evidence organization and answer construction without itself terminating execution. These intervention families describe adaptation intents rather than unique workflows. The realization map t(ι)=(ι;χt,ωt0)⊆tG_t( )=G ( ; _t, _t^0 ) _t instantiates intervention ι according to the current context, factual workflow, available tools, and task constraints. Unavailable intervention families are removed from the set considered at step t, so t(ι)≠∅G_t( )≠ for every family entering a maximization. Implementation correspondence. The intervention predictor produces a discrete intervention-family label. A workflow-construction component then converts the label into one or more structured workflow candidates. For Keep, the constructed set contains only ωt0 _t^0. For Deliberate, the harness allocates an additional reasoning step without forcing a tool or workflow change. For the three Revise modes, candidate templates constrain the next workflow toward evidence inspection, result verification, or visual analysis. For the two Stabilize modes, recent execution traces are checked for duplicate actions, repeated tool arguments, empty results, or stalled progress. AnswerSynthesis generates a nonterminal workflow that organizes available evidence into a candidate answer. Each generated workflow is parsed into the same internal representation used by the base harness before ARCO filtering and valuation. Assumption A.1: Workflow Mediation. For fixed χt _t, each ω∈t(ι)ω _t( ) is an executable realization of intervention ι , intervention ItI_t affects UtU_t only through the realized workflow WtW_t, and intervention consistency holds. This is a structural modeling restriction rather than an identification result: semantic labels index workflow families but are not assigned context-independent causal values. Under this assumption, [Ut∣do(It=ι,Wt=ω),χt]=[Ut∣do(Wt=ω),χt]. aligned E [U_t (I_t= ,W_t=ω), _t ]=E [U_t (W_t=ω), _t ]. aligned (A4) Hence, the effect associated with intervention ι is evaluated through its concrete realization ω, relative to ωt0 _t^0. The same intervention family may therefore induce different effects across contexts and workflow realizations. A.3 Error-Bounded Amortized Intervention Selection For fixed (χt,ωt0)( _t, _t^0), define the true and estimated values of intervention family ι as Vtint(ι) V_t^int( ) =maxω∈t(ι)Γt(ω), = _ω _t( ) _t(ω), (A5) V^tint(ι) V_t^int( ) =maxω∈t(ι)Γ^ϕ(ω;χt,ωt0). = _ω _t( ) _φ (ω; _t, _t^0 ). (A6) Let t=⋃ι∈ℐtwft(ι)C_t= _ _t^wfG_t( ) denote all workflows considered at step t, including the factual workflow, and define εt=supω∈t|Γ^ϕ(ω;χt,ωt0)−Γt(ω)|. _t= _ω _t | _φ (ω; _t, _t^0 )- _t(ω) |. Let ιt†=argmaxι∈ℐtwfVtint(ι) _t = _ _t^wfV_t^int( ) be a true best intervention family, ιt∗=argmaxι∈ℐtwfV^tint(ι) _t^*= _ _t^wf V_t^int( ) be the offline estimated target, and define the amortization gap ρt=V^tint(ιt∗)−V^tint(ι^t)≥0. _t= V_t^int( _t^*)- V_t^int( _t)≥ 0. Proposition A.1: Effect-and-Amortization Error Bound. For any estimated workflow-effect function and online intervention predictor, Vtint(ιt†)−Vtint(ι^t)≤2εt+ρt.V_t^int( _t )-V_t^int( _t)≤ 2 _t+ _t. (A7) Proof. For every intervention family ι , the maximum operator and the definition of εt _t imply |V^tint(ι)−Vtint(ι)|≤εt. | V_t^int( )-V_t^int( ) |≤ _t. Therefore, Vtint(ιt†) V_t^int( _t ) ≤V^tint(ιt†)+εt ≤ V_t^int( _t )+ _t ≤V^tint(ιt∗)+εt ≤ V_t^int( _t^*)+ _t =V^tint(ι^t)+ρt+εt = V_t^int( _t)+ _t+ _t ≤Vtint(ι^t)+2εt+ρt. ≤ V_t^int( _t)+2 _t+ _t. Rearranging proves Equation (A7). □ Corollary A.1: Exact Intervention Recovery. Suppose that ιt† _t is unique and define its true margin as Δtint=Vtint(ιt†)−maxι≠ιt†Vtint(ι). _t^int=V_t^int( _t )- _ ≠ _t V_t^int( ). If Δtint>2εt+ρt, _t^int>2 _t+ _t, then ι^t=ιt†. _t= _t . The result does not assume that estimated and true rankings are identical. Instead, it quantifies how workflow-effect estimation error and amortized prediction error jointly determine intervention-selection quality. When the online predictor reproduces the offline target, ρt=0 _t=0, and the selection regret is bounded by 2εt2 _t. A.4 CIEL Training and Separated Supervision CIEL separates workflow-effect supervision from factorized decision supervision. The workflow-effect estimator is trained only from offline checkpointed paired executions that compare factual and alternative workflows from the same execution context. Retrospective trajectory evidence provides weak supervision for deliberation, revision, and completion decisions, but is not treated as an identified workflow-effect observation. Checkpointed paired intervention records. For a selected training state, the harness checkpoints the execution context χt _t, including the agent context, recent execution memory, environment state, tool state, intermediate artifacts, and current task progress. The checkpoint is restored to execute the factual workflow ωt0 _t^0 and an admissible counterfactual candidate ω as two isolated branches. Both branches use the same continuation policy πcontπ^cont, evaluation horizon, model and tool configuration, task evaluator, and resource-accounting rule. Branch-specific planning, model inference, tool interaction, and execution costs are included in the resulting utility. Post-checkpoint randomness is either matched across branches or independently sampled from the same distribution and remains independent of workflow assignment. The analysis set tC_t in Appendix A.3 intentionally includes the factual workflow. For paired replay, define only the counterfactual subset tcf=t∖ωt0.C_t^cf=C_t \ _t^0\. (A8) No replay count is required for ωt0 _t^0, whose relative effect is fixed to zero. If tcf=∅C_t^cf= , no paired-effect record is constructed at that state and Keep remains the only available reference family. For each ω∈tcfω _t^cf, let Kt,ω≥1K_t,ω≥ 1 be the number of valid paired replays, and let Ut(k)(ω)andUt(k)(ωt0)U_t^(k)(ω) U_t^(k)( _t^0) denote the candidate and factual utilities in replay k. The paired workflow-effect target is Γ~tpair(ω)=1Kt,ω∑k=1Kt,ω[Ut(k)(ω)−Ut(k)(ωt0)]. _t^pair(ω)= 1K_t,ω _k=1^K_t,ω [U_t^(k)(ω)-U_t^(k)( _t^0) ]. (A9) The factual reference is fixed by construction: Γ~tpair(ωt0)=Γ^ϕ(ωt0;χt,ωt0)=0. _t^pair( _t^0)= _φ ( _t^0; _t, _t^0 )=0. Assumption A.2: State-Sufficient Paired Replay. The restored checkpoint contains the pre-intervention variables required to reproduce the execution context relevant to post-intervention performance. After restoration, the branches differ only in their assigned workflow, follow the same continuation and evaluation protocol, do not interfere, and use post-checkpoint randomness independent of workflow assignment. The candidate set, replay count, and replay-inclusion criteria are fixed before paired outcomes are observed; replay validity does not depend on the sign or magnitude of the observed utility difference. Proposition A.2: Identification and Concentration by Paired Replay. Under Assumption A.2, [Γ~tpair(ω)∣χt]=Γt(ω)E [ _t^pair(ω) _t ]= _t(ω) (A10) for every ω∈tcfω _t^cf. Suppose additionally that Ut(ω)∈[umin,umax],R=umax−umin.U_t(ω)∈[u_ ,u_ ], R=u_ -u_ . For nonempty tcfC_t^cf, define Mt=|tcf|,Ktpair=minω∈tcfKt,ω.M_t= |C_t^cf |, K_t^pair= _ω _t^cfK_t,ω. If paired replays are independent conditional on χt _t, then, with probability at least 1−ζ1-ζ, supω∈tcf|Γ~tpair(ω)−Γt(ω)|≤R2log(2Mt/ζ)Ktpair. _ω _t^cf | _t^pair(ω)- _t(ω) |≤ R 2 (2M_t/ζ)K_t^pair. (A11) Proof. Because both branches begin from the same restored context and workflow assignment is independent of post-checkpoint randomness, each branch follows the potential-outcome distribution associated with its assigned workflow. Linearity of expectation yields Equation (A10). For each candidate, the paired difference lies in [−R,R][-R,R]. Hoeffding’s inequality bounds the empirical-mean deviation, and a union bound over the MtM_t counterfactual candidates yields Equation (A11). □ Connection to intervention-selection error. Define the effect-model fitting error over paired targets as αt=supω∈tcf|Γ^ϕ(ω;χt,ωt0)−Γ~tpair(ω)|. _t= _ω _t^cf | _φ (ω; _t, _t^0 )- _t^pair(ω) |. Because the factual effect and its estimate are both zero, the error εt _t defined over the full set tC_t equals the supremum over tcfC_t^cf whenever this subset is nonempty. Therefore, with probability at least 1−ζ1-ζ, εt≤αt+R2log(2Mt/ζ)Ktpair. _t≤ _t+R 2 (2M_t/ζ)K_t^pair. Combining this inequality with Proposition A.1 gives Vtint(ιt†)−Vtint(ι^t)≤2αt+2R2log(2Mt/ζ)Ktpair+ρt. aligned V_t^int( _t )-V_t^int( _t)≤ 2 _t+2R 2 (2M_t/ζ)K_t^pair+ _t. aligned (A12) Thus, intervention-selection quality is jointly controlled by workflow-effect fitting error, finite paired-replay error, and amortized prediction error. Paired workflow-effect training set. The workflow-effect training set is pair=(χt,ωt0,ω,Γ~tpair(ω),wt,ω):ω∈tcf,D_pair= \ ( _t, _t^0,ω, _t^pair(ω),w_t,ω ):ω _t^cf \, where wt,ω≥0w_t,ω≥ 0 is a reliability weight determined from pre-specified replay-validity diagnostics, replay count, and comparison stability, without using the sign of the observed effect. Pairs are excluded when the checkpoint cannot be restored, the branches do not share the same continuation protocol, or branch outcomes cannot be isolated. Only checkpointed paired executions are treated as workflow-effect observations. Unpaired trajectories, incomplete comparisons, and model-scored candidates do not directly supervise Γ^ϕ _φ and are not interpreted as identified causal effects. The workflow-effect estimator is trained by ℒeff=(χt,ωt0,ω,Γ~tpair(ω),wt,ω)∼pair[wt,ω(Γ^ϕ(ω;χt,ωt0)−Γ~tpair(ω))2]. aligned L_eff=E_ ( _t, _t^0,ω, _t^pair(ω),w_t,ω ) _pair [w_t,ω ( _φ (ω; _t, _t^0 )- _t^pair(ω) )^2 ]. aligned (A13) Intervention target and objective. Candidate workflows are grouped by semantic intervention family. The offline target is induced by the learned workflow-effect estimator: ιt∗=argmaxι∈ℐtwfmaxω∈t(ι)Γ^ϕ(ω;χt,ωt0). _t^*= _ _t^wf _ω _t( ) _φ (ω; _t, _t^0 ). (A14) When no alternative family has a positive estimated effect, conservative tie-breaking assigns ιt∗=Keep. _t^*= Keep. The amortized intervention predictor is trained by ℒint=−logpθu(ιt∗∣χt,ωt0).L_int=- p_ _u ( _t^* _t, _t^0 ). (A15) Trajectory evidence for factorized decisions. For each intervention step, the implementation extracts ℰtdec=(Yτ,Yref,Eterr,Ctmain,Ctplan,rt,ct,Δtcf,ℋtexec),E_t^dec= (Y_τ,Y_ref,E_t^err,C_t^main,C_t^plan,r_t,c_t, _t^cf,H_t^exec ), where YτY_τ and YrefY_ref are the current and reference task outcomes; EterrE_t^err records execution failure; CtmainC_t^main and CtplanC_t^plan record main-agent and planner costs; rtr_t and ctc_t summarize route and candidate statistics; and ℋtexecH_t^exec contains progress, repetition, failure, and completion evidence. A decision-target operator produces Ψdec(ℰtdec)=(zt∗,ytchg,∗,bt∗). _dec (E_t^dec )= (z_t^*,y_t^chg,*,b_t^* ). These are outcome-conditioned weak decision labels, not observations of Γt(ω) _t(ω). Counterfactual Deliberation Effect target. The target zt∗∈Skip,Light,Fullz_t^*∈ \ Skip, Light, Full \ specifies the preferred amount of additional deliberation. Skip is assigned when the factual workflow is sufficient or additional planning produces no useful decision evidence. Light is assigned when restricted diagnostics or a small candidate set is sufficient. Full is assigned only when broader generation or evaluation materially improves intervention selection, recovers ineffective execution, or prevents a high-risk factual action. Among routes preserving decision quality and task success, the target selects the least costly route. Counterfactual Revision Effect target. The target ytchg,∗∈Keep,Changey_t^chg,*∈ \ Keep, Change \ specifies whether the selected candidate ωt+ _t^+ should replace the factual workflow. Change requires evidence that the candidate improves the factual execution while preserving success and satisfying authorization conditions. When paired execution evidence is available, its observed workflow difference provides the primary comparison. Unpaired retrospective evidence may provide auxiliary decision supervision but does not enter pairD_pair. Unsafe, invalid, task-incompatible, over-budget, duplicate, ineffective, or success-breaking candidates receive Keep supervision. Completion Attribution Effect target. The target bt∗∈0,1b_t^*∈\0,1\ indicates whether task completion is supported by evidence already available at step t. The label bt∗=1b_t^*=1 requires retrospective evidence that the task requirements were already satisfied, such as a valid final answer, passing tests, a required artifact, or an environment-provided success state. Budget exhaustion, inactivity, repeated failure, predicted futility, or the absence of another action does not produce a positive completion target. Head-specific objectives. The factorized decision heads are trained by ℒCDE _CDE =−logpθp(zt∗∣χt,ωt0,ι^t), =- p_ _p (z_t^* _t, _t^0, _t ), (A16) ℒCRE _CRE =−logpθc(ytchg,∗∣χt,ωt0,ωt+,ι^t), =- p_ _c (y_t^chg,* _t, _t^0, _t^+, _t ), (A17) ℒCAE _CAE =−logpθe(bt∗∣χt,ot,ht). =- p_ _e (b_t^* _t,o_t,h_t ). (A18) The complete CIEL objective is ℒCIEL=ℒeff+ℒint+ℒCDE+ℒCRE+ℒCAE.L_CIEL=L_eff+L_int+L_CDE+L_CRE+L_CAE. (A19) The objectives use distinct sources of supervision: ℒeffL_eff learns workflow effects from offline checkpointed paired executions; ℒintL_int amortizes the preference induced by the learned workflow effects; and ℒCDEL_CDE, ℒCREL_CRE, and ℒCAEL_CAE learn route, revision, and completion decisions from success-preserving trajectory evidence. This separation prevents heuristic or model-scored trajectory signals from being interpreted as identified workflow-effect observations. Appendix B: ARCO Realization and Authorization Details This appendix explains how Advantage-Realizing Causal Orchestration (ARCO) implements CIEL decisions through route selection, candidate generation, causal–operational valuation, workflow authorization, and completion verification. B.1 Route Feasibility and Candidate Construction Let =Skip,Light,Full,Skip≺Light≺Full.Z= \ Skip, Light, Full \, Skip Light Full. The ordering represents increasing deliberation depth, candidate-generation coverage, and execution cost. The provisional route is z~t=argmaxz∈pθp(z∣χt,ωt0,ι^t). z_t= _z p_ _p (z _t, _t^0, _t ). (B1) Let Cplan(z)C^plan(z) be the predicted planning cost of route z, BtplanB_t^plan the remaining planning budget, NtfullN_t^full the number of previous full-deliberation calls, BfullB^full their maximum allowed number, tlastfullt_last^full the most recent full-deliberation step, κ the full-route cooldown, (z)T(z) the tools or capabilities required by route z, and tavailT_t^avail the currently available tools and capabilities. If no full route has previously been executed, set tlastfull=−∞t_last^full=-∞. Define the common planning-and-tool feasibility indicator as gt(z)=[Cplan(z)≤Btplan][(z)⊆tavail].g_t(z)=I [C^plan(z)≤ B_t^plan ]I [T(z) _t^avail ]. (B2) The route feasibility is ft(z)=1,z=Skip,gt(z),z=Light,gt(z)[Ntfull<Bfull,t−tlastfull≥κ],z=Full.f_t(z)= cases1,&z= Skip,\\ g_t(z),&z= Light,\\ g_t(z)\,I [ array[]cN_t^full<B^full,\\[-2.84526pt] t-t_last^full≥κ array ],&z= Full. cases (B3) The feasible route set is ℱt=z∈:ft(z)=1.F_t= \z :f_t(z)=1 \. Because Skip is always feasible, ℱt≠∅F_t≠ . The executed route is zt=max≺z∈ℱt:z⪯z~t.z_t= _ \z _t:z z_t \. (B4) Thus, an infeasible prediction is downgraded to the highest feasible route and is never upgraded beyond the learned prediction. The raw route-specific candidate set is Ωtraw=ωt0,zt=Skip,ωt0∪tlight(ι^t),zt=Light,ωt0∪t(ι^t),zt=Full. _t^raw= cases\ _t^0\,&z_t= Skip,\\[2.84526pt] \ _t^0\ _t^light ( _t ),&z_t= Light,\\[2.84526pt] \ _t^0\ _t ( _t ),&z_t= Full. cases (B5) Here, tlight(ι^t)⊆t(ι^t)G_t^light( _t) _t( _t) is a restricted, low-cost realization set. Candidate exposure is bounded by KSkip=1,KLight<KFull.K_ Skip=1, K_ Light<K_ Full. These route budgets are distinct from the paired-replay quantity KtpairK_t^pair in Appendix A.4. Let S^t(ω∣χt) S_t(ω _t) be the workflow-safety score and ξt _t its minimum accepted threshold. Before expensive valuation, the retained candidate set is Ωt=ωt0∪TopKKzt−1ω∈Ωtraw∖ωt0:S^t(ω∣χt)≥ξt, aligned _t=\ _t^0\ _K_z_t-1 \ω∈ _t^raw \ _t^0\: S_t(ω _t)≥ _t \, aligned (B6) where TopK0(⋅)=∅TopK_0(·)= . When more candidates pass the safety threshold than the route budget allows, TopKTopK retains them according to the low-cost workflow-construction score produced before full valuation. The admissible route-exposed set is Ω¯t=Ωt∩t. _t= _t _t. (B7) Because ωt0∈t _t^0 _t, the filtered set is nonempty. Implementation correspondence. The CDE head outputs scores for the three route classes. A deterministic feasibility stage checks route-specific planning budgets, full-route limits and cooldown, required tools, and environment capabilities. Under Skip, planner and candidate-generation calls are bypassed. Under Light, the implementation uses restricted diagnostics and a small candidate budget. Under Full, all configured route-compatible generation and evaluation operations are available. The implementation records the provisional route, executed route, downgrade reason, exposed candidates, prescreening decisions, and incurred planning cost. B.2 Causal–Operational Valuation For each ω∈Ω¯tω∈ _t, let P^t(ω) P_t(ω), C^t(ω) C_t(ω), R^t(ω) R_t(ω), I^t(ω) I_t(ω), B^t(ω) B_t(ω), and S^t(ω∣χt) S_t(ω _t) denote estimated task progress, execution cost, operational risk, information gain, robustness, and workflow safety, respectively. All available signals are calibrated to a common bounded range before aggregation. A signal unavailable for a benchmark is omitted rather than assigned an arbitrary value. Define the shared progress–risk–safety score and the full-route information–robustness bonus as Q^t(ω) Q_t(ω) =αprogP^t(ω)−αriskR^t(ω)+αsafeS^t(ω∣χt), = _prog P_t(ω)- _risk R_t(ω)+ _safe S_t(ω _t), (B8) H^t(ω) H_t(ω) =αinfoI^t(ω)+αrobB^t(ω). = _info I_t(ω)+ _rob B_t(ω). (B9) The route-conditioned operational utility is U^op(ω∣χt,zt)=0,zt=Skip,Q^t(ω)−αcostlightC^t(ω),zt=Light,Q^t(ω)−αcostfullC^t(ω)+H^t(ω),zt=Full. aligned U_op (ω _t,z_t )= cases0,&z_t= Skip,\\ Q_t(ω)- _cost^light C_t(ω),&z_t= Light,\\ Q_t(ω)- _cost^full C_t(ω)+ H_t(ω),&z_t= Full. cases aligned (B10) All coefficients are nonnegative. The light route omits information-gain and robustness estimation when restricted diagnostics are sufficient, whereas the full route uses the complete configured valuation. The causal–operational utility is U^ARCO(ω∣χt,zt,Γ^ϕ,t)=U^op(ω∣χt,zt)+ηΓΓ^ϕ,t(ω), U_ARCO (ω _t,z_t, _φ,t )= U_op (ω _t,z_t )+ _ _φ,t(ω), (B11) where ηΓ≥0 _ ≥ 0 controls the contribution of estimated workflow effect evidence. The strongest exposed candidate is ωt+=argmaxω∈Ω¯tU^ARCO(ω∣χt,zt,Γ^ϕ,t). _t^+= _ω∈ _t U_ARCO (ω _t,z_t, _φ,t ). (B12) Its advantage over the factual workflow is Δtcf= _t^cf= U^op(ωt+∣χt,zt)−U^op(ωt0∣χt,zt) U_op ( _t^+ _t,z_t )- U_op ( _t^0 _t,z_t ) (B13) +ηΓ[Γ^ϕ,t(ωt+)−Γ^ϕ,t(ωt0)]. + _ [ _φ,t( _t^+)- _φ,t( _t^0) ]. The estimator enforces Γ^ϕ,t(ωt0)=0. _φ,t( _t^0)=0. Implementation correspondence. Task-progress signals may be derived from newly acquired evidence, intermediate environment states, test results, or predicted goal advancement. Cost signals use expected or observed model tokens, planner calls, tool calls, and latency. Risk signals capture invalid actions, destructive operations, execution errors, or uncertain state changes. Information-gain signals estimate whether the candidate obtains evidence absent from the factual path. Robustness signals favor candidates that depend less on unverified assumptions or fragile execution sequences. The final score ranks only candidates in the route-exposed admissible set Ω¯t _t. B.3 Conservative Workflow Authorization The CRE decision for the selected candidate is ytchg=argmaxy∈Keep,Changepθc(y∣χt,ωt0,ωt+,ι^t).y_t^chg= _y∈\ Keep, Change\p_ _c (y _t, _t^0, _t^+, _t ). (B14) Keep retains the factual workflow, whereas Change provides learned support for replacement by ωt+ _t^+. CRE and causal–operational valuation have complementary roles: CRE determines whether replacement is contextually supported, while Δtcf _t^cf determines whether the estimated improvement is large enough to justify revision. The authorization rule is ωt⋆=ωt+,ytchg=Change∧Δtcf≥δι^t,ωt0,otherwise, _t = cases _t^+,&y_t^chg= Change\ \ _t^cf≥ _ _t,\\[2.84526pt] _t^0,&otherwise, cases (B15) where δι^t≥0 _ _t≥ 0 is the authorization margin for the predicted intervention family. Because ωt+ _t^+ is selected from Ω¯t _t, admissibility, safety, schema, task-compatibility, and budget checks have already been satisfied before this rule is applied. For implementation diagnostics, define atsup=[ytchg=Change][Δtcf≥δι^t],atfeas=[ωt+∈Ω¯t], aligned a_t^sup=I [y_t^chg= Change ]I [ _t^cf≥ _ _t ],a_t^feas=I [ _t^+∈ _t ], aligned (B16) and atchg=atsupatfeas.a_t^chg=a_t^supa_t^feas. The learned CRE output and final authorization result are logged separately, allowing prediction errors to be distinguished from constraint-triggered fallback. If either support or feasibility fails, ARCO retains ωt0 _t^0. Scope of workflow optimality. ARCO does not claim global optimization over the complete admissible workflow space tA_t. Instead, ωt+=argmaxω∈Ω¯tU^ARCO(ω∣χt,zt,Γ^ϕ,t) _t^+= _ω∈ _t U_ARCO (ω _t,z_t, _φ,t ) is optimal only within the route-exposed admissible set. CHILL-Harness executes this candidate when authorized and otherwise retains the factual workflow. B.4 Evidence-Grounded Completion and Answer Synthesis The CAE prediction is b^t=argmaxb∈0,1pθe(b∣χt,ot,ht). b_t= _b∈\0,1\p_ _e (b _t,o_t,h_t ). (B17) The value b^t=1 b_t=1 indicates learned completion support, whereas b^t=0 b_t=0 indicates that completion is unsupported by the CAE head. An explicit task verifier produces etcomp=[t(χt,ot,ht)=1],e_t^comp=I [V_t ( _t,o_t,h_t )=1 ], (B18) where tV_t is the task-specific completion verifier. Termination is authorized by dtterm=b^tetcomp.d_t^term= b_te_t^comp. (B19) Thus, termination requires both learned completion support and explicit task evidence. The verifier may decompose completion evidence into strong, weak, and contradictory signals: etcomp=[Etstrong=1∨(Etweak=1∧Etfail=0)].e_t^comp=I [E_t^strong=1\ \ (E_t^weak=1\ \ E_t^fail=0 ) ]. Strong evidence includes passing tests, successful builds, accepted solutions, or explicit environment success. Weak evidence includes a valid final answer or task-completion signal, whereas traceback, assertion, permission, and execution failures constitute contradictory evidence. Implementation correspondence. The CAE head does not directly invoke the terminal action; its output is passed to the task-specific verifier. Depending on the benchmark, the verifier checks a required answer format, successful tests or evaluator results, the existence and validity of a required artifact, an environment-provided success state, or another explicit completion condition. Predicted futility, budget exhaustion, repeated failure, or lack of progress does not set etcomp=1e_t^comp=1. Such signals may instead redirect execution toward AnswerSynthesis or another workflow adaptation. Therefore, AnswerSynthesis≠Complete. AnswerSynthesis≠ Complete. The former is a nonterminal adaptation; the latter is an authorized terminal decision. The implementation records the CAE output, verifier result, supporting evidence, synthesis decision, and final terminal action. Appendix C: Success-Preserving Learning and Complete Inference This appendix specifies the trajectory-level objective, the separated supervision scheme, local authorization constraints, offline paired-effect collection, and the complete CHILL-Harness inference procedure. C.1 Success-Preserving Trajectory Objective The learnable CIEL parameter set is Θ=ϕ,θu,θp,θc,θe. = \φ, _u, _p, _c, _e \. ARCO introduces no separate prediction parameters in the current formulation. It realizes CIEL outputs through candidate generation, valuation, filtering, and rule-based authorization with calibrated weights and thresholds. Let K denote the resource channels measured for a benchmark. The trajectory-level execution cost is eff(τ)=∑k∈λkCk(τ),J_eff(τ)= _k _kC_k(τ), (C1) where Ck(τ)C_k(τ) is the amount of resource k consumed by trajectory τ, and λk≥0 _k≥ 0 is its weight. For the reported experiments, the principal resource channels are Ctok(τ)=Cmainin(τ)+Cmainout(τ)+Cplanin(τ)+Cplanout(τ),C_tok(τ)=C_main^in(τ)+C_main^out(τ)+C_plan^in(τ)+C_plan^out(τ), and Ctime(τ)=Tmain(τ)+Tplan(τ)+Ttool(τ)+Tenv(τ)+Tverify(τ). aligned C_time(τ)=T_main(τ)+T_plan(τ)+T_tool(τ)+T_env(τ)+T_verify(τ). aligned (C2) The reported total token count includes main-agent and counterfactual-planner calls made during benchmark inference. Offline paired-replay collection is a training-data construction procedure and is not included in test-time token or runtime measurements; its cost is tracked separately. The global objective is minΘ _ τ∼πCHILL,Θ[eff(τ)] _τ _CHILL, [J_eff(τ) ] (C3) s.t. PassRate(πCHILL,Θ)≥PassRate(πref)−ϵpass. ( _CHILL, ) ( _ref )- _pass. Here, πCHILL,Θ _CHILL, is the execution policy induced by CIEL and ARCO; πref _ref is the matched reference harness; PassRate(π)PassRate(π) is the fraction of evaluated tasks completed successfully by policy π; and ϵpass≥0 _pass≥ 0 is the allowed aggregate pass-rate degradation. Success is determined by the benchmark evaluator rather than by the harness completion prediction. C.2 Trajectory Supervision and Decision-Target Construction Let Yτ∈0,1andYref∈0,1Y_τ∈\0,1\ Y_ref∈\0,1\ indicate whether the CHILL-Harness and matched reference trajectories succeed, respectively. The trajectory supervision score is ℒtraj(τ)=eff(τ)+λfail(1−Yτ)+λbreak[Yref=1∧Yτ=0]. aligned L_traj(τ)=J_eff(τ)+ _fail (1-Y_τ )+ _breakI [Y_ref=1\ \ Y_τ=0 ]. aligned (C4) The first term measures execution cost, the second penalizes task failure, and the third penalizes adaptations that break a successful reference trajectory. The coefficients λfail,λbreak≥0 _fail, _break≥ 0 control the two failure penalties. Separated supervision correspondence. The trajectory score is not differentiated through external tools or the environment. Instead, it determines valid samples, confidence weights, and weak decision targets: ℒtraj⟶(zt∗,ytchg,∗,bt∗).L_traj (z_t^*,y_t^chg,*,b_t^* ). Workflow-effect supervision is constructed separately: pair⟶Γ~tpair(ω)⟶Γ^ϕ⟶ιt∗.D_pair _t^pair(ω) _φ _t^*. Thus, retrospective trajectory evidence does not directly define or supervise the identified workflow-effect target. CDE targets select the least costly route preserving decision quality; CRE targets determine whether the selected candidate should replace the factual workflow; and CAE targets determine whether completion is already supported by evidence. C.3 Local Authorization Constraints Using the unified safety score from Appendix B, the admissible workflow set is t=ω: _t=\ω: S^t(ω∣χt)≥ξt S_t (ω _t )≥ _t (C5) ∧SchemaValid(ω) \ SchemaValid(ω) ∧TaskCompatible(ω;χt) \ TaskCompatible (ω; _t ) ∧WithinBudget(ω;χt) \ WithinBudget (ω; _t ) ∧¬Repeatedt(ω). \ _t(ω)\. Here, S^t(ω∣χt) S_t(ω _t) is the workflow-safety score and ξt _t its minimum threshold; SchemaValidSchemaValid checks workflow and tool-call structure; TaskCompatibleTaskCompatible checks that proposed operations are permitted in the current task and environment; WithinBudgetWithinBudget checks remaining model, planning, tool, revision, and execution budgets; and RepeatedtRepeated_t identifies duplicate or repeatedly ineffective workflows. Each generated candidate is parsed and passed through these checks before valuation and authorization. Rejected candidates are removed from Ω¯t=Ωt∩t, _t= _t _t, and their rejection reasons are stored in the execution trace. Revision satisfies ωt⋆=ωt+⟹[ytchg=Change∧Δtcf≥δι^t], _t = _t^+ [y_t^chg= Change\ \ _t^cf≥ _ _t ], and completion satisfies dtterm=1⟹[b^t=1∧etcomp=1].d_t^term=1 [ b_t=1\ \ e_t^comp=1 ]. Together, these constraints ensure ωt⋆∈Ω¯t⊆t _t ∈ _t _t and prevent unsupported revision or premature termination. C.4 Complete Inference, Offline Pairing, and Logging Algorithm 2 Complete CHILL-Harness Inference 1: Input: χt,ωt0,ot,ht _t, _t^0,o_t,h_t; Output: ωt⋆,dtterm _t ,d_t^term 2: Predict ι^t=argmaxι∈ℐtwfpθu(ι∣χt,ωt0) _t= _ _t^wfp_ _u ( _t, _t^0 ) 3: Predict the provisional route z~t=argmaxz∈pθp(z∣χt,ωt0,ι^t) z_t= _z p_ _p (z _t, _t^0, _t ) and select the highest feasible route zt⪯z~tz_t z_t 4: Generate the raw route-conditioned candidate set Ωtraw _t^raw using Equation (B5) 5: Prescreen candidates using Equation (B6) and obtain Ωt _t 6: Filter admissible candidates: Ω¯t=Ωt∩t _t= _t _t 7: For each ω∈Ω¯tω∈ _t, estimate Γ^ϕ,t(ω)=Γ^ϕ(ω;χt,ωt0) _φ,t(ω)= _φ (ω; _t, _t^0 ) and compute U^ARCO(ω∣χt,zt,Γ^ϕ,t) U_ARCO(ω _t,z_t, _φ,t) 8: Select ωt+ _t^+ and compute Δtcf _t^cf using Equations (B12) and (B13) 9: Predict CRE support: ytchg=argmaxy∈Keep,Changepθc(y∣χt,ωt0,ωt+,ι^t)y_t^chg= _y∈\ Keep, Change\p_ _c (y _t, _t^0, _t^+, _t ) 10: Authorize ωt⋆ _t using Equation (B15) 11: Compute b^t=argmaxb∈0,1pθe(b∣χt,ot,ht) b_t= _b∈\0,1\p_ _e (b _t,o_t,h_t ) and dtterm=b^t[t(χt,ot,ht)=1]d_t^term= b_tI [V_t ( _t,o_t,h_t )=1 ] 12: if dtterm=1d_t^term=1 then 13: Invoke the task-terminal action 14: else 15: Execute ωt⋆ _t 16: end if 17: Log decisions, candidate evaluations, rejection reasons, execution costs, observations, errors, and task outcomes The inference record contains at least (χt,ωt0,ι^t,z~t,zt,Ωtraw,Ωt,Ω¯t,ωt+,Δtcf,ytchg,ωt⋆,b^t,etcomp,dtterm), aligned ( _t, _t^0, _t, z_t,z_t, _t^raw, _t, _t, _t^+, _t^cf,y_t^chg, _t , b_t,e_t^comp,d_t^term ), aligned (C6) together with tool calls, observations, rejection reasons, costs, errors, and final task outcomes. Offline paired-effect collection. Checkpointed paired replay is performed only during offline training-data construction. For each selected training checkpoint and counterfactual candidate, the implementation restores the same checkpoint, executes the factual and candidate branches under the common continuation and evaluation protocol, computes Equation (A9), and adds the resulting record to pairD_pair. CIEL parameters are then trained on the designated training split and frozen before benchmark evaluation. No paired replay, effect-model update, or decision-head update is performed on evaluation tasks. The complete learning and execution flow is offline paired execution⟶Γ~tpair⟶Γ^ϕ⟶ιt∗,offline paired execution _t^pair _φ _t^*, trajectory evidence⟶(zt∗,ytchg,∗,bt∗),trajectory evidence (z_t^*,y_t^chg,*,b_t^* ), frozen CIEL predictions⟶ARCO realization⟶authorized execution.frozen CIEL predictions realization execution.