Paper deep dive

The Handoff Tax: Continuing Non-Native Trajectories in LLM Agents

Roy Ganz, Mor Shpigel Nacson, Adi Kalyanpur, Ron Litman

Year: 2026Venue: arXiv preprintArea: cs.AIType: PreprintEmbeddings: 79

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 94%

Last extracted: 8/29/2026, 4:32:58 AM

Summary

This paper investigates the 'Handoff Tax' in long-running LLM coding agents, analyzing the cost-quality trade-offs when switching between low-cost/low-capability (LC) and high-cost/high-capability (HC) models. The study evaluates escalation (LC to HC) and downshift (HC to LC) across different handoff interfaces (Raw, Compaction, Trajectory Removal). Key findings indicate that raw escalation recovers less than half the quality gap while incurring significant cost premiums, whereas downshift offers a favorable cost-quality balance. The value of inherited trajectory information is direction-dependent: reducing LC trajectory information improves escalation quality, while preserving HC trajectory information is crucial for maintaining downshift quality.

Entities (14)

SWE-bench Verified · dataset · 99%Handoff Tax · concept · 98%Claude · model-family · 97%GPT · model-family · 97%Downshift · process · 96%escalation · process · 96%GPT-5.6 Luna · model · 95%GPT 5.6 Sol · model · 95%Haiku 4.5 · model · 95%Opus 4.7 · model · 95%Compactpre · interface · 94%Compactsuf · interface · 94%Raw Handoff · interface · 94%Traj-drop · interface · 94%

Relation Signals (11)

Haiku 4.5 → ispartof → Claude

confidence 98% · Haiku 4.5 ... for Claude

GPT 5.6 Sol → ispartof → GPT

confidence 98% · GPT-5.6 Sol for GPT

GPT-5.6 Luna → ispartof → GPT

confidence 98% · GPT-5.6 Luna ... for GPT

Opus 4.7 → ispartof → Claude

confidence 98% · Opus 4.7 ... for Claude

Handoff Tax → describes → Cost-Quality Penalty

confidence 95% · We term this cost–quality penalty the handoff tax.

SWE-bench Verified → usedforevaluation → Handoff Tax Study

confidence 95% · We study these questions systematically on SWE-bench Verified

Downshift → offers → Favorable cost-quality point

confidence 93% · By contrast, downshift offers a favorable cost-quality point.

Raw Escalation → incurs → Substantial cost premium

confidence 92% · incurring a substantial cost premium

Raw Escalation → recovers → Less than half of HC quality gap

confidence 92% · Raw escalation recovers less than half of the LC-to-HC quality gap

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Coding agents perform long-running tasks spanning dozens of model calls, tool uses, and code edits. As these runs unfold, users face a practical cost-quality trade-off: escalating to a stronger model when a cheaper one struggles, or downshifting once the hard reasoning is complete. Each switch requires the receiver to continue a non-native trajectory produced by another model. We study how this handoff affects quality and cost, and how varying the trajectory information inherited by the receiver changes the outcome. Using pairs of low-cost, low-capability (LC) and high-cost, high-capability (HC) models from the Claude and GPT families, we vary handoff direction, timing, and interface, comparing full-trajectory transfer, compaction, and trajectory removal while preserving the repository state. Across both model families, full-trajectory escalation recovers less than half of the LC-to-HC quality gap while incurring a substantial cost premium. We term this cost-quality penalty the handoff tax. By contrast, downshift offers a favorable cost-quality point. Interestingly, the preferred interface also reverses with direction: reducing LC-model trajectory information improves escalation quality, whereas removing the HC-model trajectory reduces downshift quality.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

75,773 characters extracted from source content.

Expand or collapse full text

The Handoff Tax: Continuing Non-Native Trajectories in LLM Agents Roy Ganz †thanks: Equal contribution.†thanks: Correspondence: royganz@amazon.com. Mor Shpigel Nacson* Adi Kalyanpur Ron Litman Affiliation: AWS, Agentic AI Abstract Coding agents perform long-running tasks spanning dozens of model calls, tool uses, and code edits. As these runs unfold, users face a practical cost–quality trade-off: escalating to a stronger model when a cheaper one struggles, or downshifting once the hard reasoning is complete. Each switch requires the receiver to continue a non-native trajectory produced by another model. We study how this handoff affects quality and cost, and how varying the trajectory information inherited by the receiver changes the outcome. Using pairs of low-cost, low-capability (LC) and high-cost, high-capability (HC) models from the Claude and GPT families, we vary handoff direction, timing, and interface, comparing full-trajectory transfer, compaction, and trajectory removal while preserving the repository state. Across both model families, full-trajectory escalation recovers less than half of the LC-to-HC quality gap while incurring a substantial cost premium. We term this cost–quality penalty the handoff tax. By contrast, downshift offers a favorable cost–quality point. Interestingly, the preferred interface also reverses with direction: reducing LC-model trajectory information improves escalation quality, whereas removing the HC-model trajectory reduces downshift quality. 1 Introduction Figure 1: The Handoff Tax. (a) Users escalate to a high cost model (HC) when the cheaper model (LC) struggles, or downshift to a cheaper one once hard reasoning appears complete. (b) At handoff, the receiver inherits another model’s trajectory. We ask how this affects cost and quality, and how varying the transferred trajectory information shapes the effect. Coding agents now routinely chain dozens to hundreds of model calls on a single task, making model choice an economic decision as much as a technical one Yao et al. (2023); Schick et al. (2023); Qin et al. (2023); Patil et al. (2023); Wang et al. (2024); Wang et al. (2025); Xu et al. (2025); Hitzig et al. (2026). Available models span distinct cost–quality profiles: higher-capability models resolve more issues but are substantially more expensive, whereas cheaper models offer lower cost at reduced capability. Users must often choose among these models before knowing what level of capability the task actually requires. Model switching offers a natural response and is already supported directly by coding-agent products. Users may start with a low-capability model and escalate to a stronger one when the agent struggles; conversely, once the hard reasoning appears complete, they may downshift to a cheaper model to finish the task at lower cost. Throughout, LC and HC denote the lower-cost/lower-capability and higher-cost/higher-capability models within a family; in our experiments, these correspond to Claude and GPT model pairs. Despite being exposed in several coding-agent interfaces (e.g., /model commands in Kiro, Codex, and Claude Code), the cost–quality consequences of capability handoffs in long-horizon coding agents remain poorly understood. A mid-task handoff places a model in an unusual position: it must continue a long trajectory that another model produced. Unlike ordinary generation, where models extend their own trajectory–their own phrasing, hypotheses, tool-use idioms, and dead ends–a handoff requires them to inherit a trajectory they did not create. This inherited context may contain reasoning the receiver would not have generated and mistakes it would not have made. Whether it helps or hurts may depend on direction: an HC model might be anchored by LC’s wrong turns, while an LC model might coast on HC’s groundwork–or flounder when extending reasoning beyond its own ability. We therefore ask two questions: whether and how inheriting a non-native trajectory changes the cost–quality trade-off, and how the information transferred at the handoff shapes that effect. We study these questions systematically on SWE-bench Verified Jimenez et al. (2024) along three axes: the direction of the switch (LC→ and HC→ ), its timing (swept across difficulty-calibrated percentiles), and the interface, which determines what trajectory information the receiver inherits. The standard raw handoff passes the full trajectory verbatim. Alternatively, the trajectory can be compacted into a summary written by either the outgoing or incoming model, or dropped while preserving the code edits already written to disk (Traj-drop). Fig. 2 illustrates these interfaces. Findings and contributions. We present, to our knowledge, the first systematic study of mid-trajectory model capability handoffs in long-horizon coding agents. Across two model families, both handoff directions, seven switch points, and four interfaces, our primary study comprises 58 configurations per family, 58,000 runs, and 36 billion tokens processed overall. Three main findings emerge from the coding-agent study: 1. Raw escalation recovers limited quality and can be dominated by restarting with HC. Raw continuation, the default escalation interface, recovers less than half of HC’s quality advantage in both model families. Most strikingly, for Claude, even after paying for the LC prefix, abandoning the attempt and restarting with HC is cheaper and more accurate than Raw continuation. 2. Continuing another model’s trajectory has direction-dependent value. Unlike escalation, Raw downshift provides a favorable cost–quality trade-off: LC receivers retain substantial quality gains over LC while remaining cheaper than HC. The balance differs across families: Claude retains most of LC’s cost advantage, whereas GPT retains most of HC’s quality advantage. 3. Handoff interfaces reveal what trajectory information helps the receiver. In escalation, reducing LC trajectory information improves the trade-off: compaction favors savings, whereas trajectory removal favors quality recovery. In downshift, removing the HC trajectory while preserving its working-tree edits reduces quality across both model families. This reveals a directional duality: HC trajectories guide LC receivers, whereas LC trajectories burden HC receivers. 2 Related Work Figure 2: Handoff interfaces. All strategies preserve the prefix model’s edited working tree at the switch but vary the trajectory information transferred to the suffix: the full trajectory (Raw), a summary written by the prefix or suffix model (Compactpre/Compactsuf), or none (Traj-drop). Prior work improves the cost–quality trade-off of LLM inference through model selection. Cascades invoke higher-capability models when lower-cost ones are insufficient, while routers dispatch requests over a model pool (Chen et al., 2023; Yue et al., 2024; Hu et al., 2024; Ding et al., 2024; Aggarwal et al., 2025; Valkanas et al., 2025; Ong et al., 2025; Kotte, 2026). Recent work extends routing to multi-turn interactions, deciding which model acts and when (Zhang et al., 2026a; Zhang et al., 2026b; Hemadri et al., 2026). SWE-Router (Son et al., 2026) uses an LC model’s partial coding trajectory to decide whether it should continue or an HC model should restart, but does not transfer the trajectory. Khraishi et al. (2026) study directional drift when one model continues another’s dialogue prefix, but only at the final turn and without alternative handoff interfaces. KC and Budathoki (2026) study agents taking over interrupted coding tasks, examining how the information exposed from the prior run affects the effort needed to rediscover the predecessor’s work. Their models are evaluated as successors rather than as LC–HC alternatives with distinct cost profiles. Our work instead studies a practical user pattern: switching during an active run between commercial models with different capability and cost profiles. We vary direction, timing, and interface to measure how continuing a non-native trajectory affects end-to-end quality and monetary cost. 3 Experimental Framework We study mid-trajectory model handoffs in long-horizon coding tasks, using SWE-bench Verified Jimenez et al. (2024), a real GitHub issue-resolution benchmark with executable pass/fail evaluation and difficulty annotations. All conditions use the same mini-swe-agent11 1 https://github.com/swe-agent/mini-swe-agent scaffold, tools, and prompts Yang et al. (2024). Within each family, LC denotes the lower-cost, lower-capability model and HC the higher-cost, higher-capability model: Haiku 4.5 Anthropic (2025) and Opus 4.7 Anthropic (2026) for Claude, and GPT-5.6 Luna and Sol for GPT OpenAI (2026). 3.1 Handoff Conditions Directions. We evaluate the resulting cost–quality trade-off in both model switch directions. Escalation (LC→ ) starts with LC and switches to HC. It tests whether HC can rescue a trajectory begun by LC and in what cost. Downshift (HC→ ) starts with HC and transfers the remaining work to LC. It tests how much of HC’s quality advantage remains and how it affects the cost. Handoff Strategies. The central design variable is what crosses the model boundary. Let 1:KT_1:K denote the trajectory produced by the prefix model up to step K, and KW_K the working tree (edited files on disk) at that point. All strategies preserve KW_K; they differ only in the trajectory information passed to the suffix model (Figure 2): • Raw: the full trajectory 1:KT_1:K is transferred. • Compactpre: the prefix model compacts 1:KT_1:K; only the summary is passed. • Compactsuf: the suffix model compacts 1:KT_1:K and continues from that summary alone. • Traj-drop: no trajectory information is transferred; the suffix model begins with only KW_K. For escalation, we construct two restart controls that charge for LC’s work but discard both its trajectory and working-tree edits before restarting HC from the original task state. Abort + HC fresh charges for LC through step K, whereas LC-full + HC-full charges for a complete LC run. These test whether any continuation of the LC trajectory is preferable to abandoning it outright. 3.2 Switch Points A handoff’s effect depends on when it occurs. Fixed step counts are not comparable across tasks or models because trajectory lengths vary. We therefore define switch points as percentiles of the starting model’s step-count distribution. Computing these percentiles over the full dataset would skew the difficulty distribution at each switch point relative to the full dataset, since easier tasks tend to finish earlier than harder ones. To avoid this bias, we estimate each percentile separately within each task-difficulty bucket. For example, p25p25 is the 25th-percentile step count for the relevant starting model and difficulty bucket. We sweep K over percentiles 5,10,15,25,35,45,50\5,10,15,25,35,45,50\. The main results average over all seven switch points. We omit later percentiles because too few instances reach the switch for reliable comparison. Appendix Tab. 11 report the exact bucket-specific steps. 3.3 Evaluation on the Switched Subset Comparing handoffs fairly requires accounting for which runs actually underwent a handoff. If the starting model finishes before step K, no switch occurs; the run is a single-model trajectory and provides no evidence about the handoff effect. We therefore evaluate the switched subset: for each direction and K, we restrict comparisons to the intersection of instances that switched under all four strategies. This ensures that every strategy is evaluated on the same tasks that underwent a handoff. We evaluate the baselines on this same subset to provide matched quality and cost references. 3.4 Metrics We report raw pass rate, mean cost (USD), and mean step count. Costs follow provider pricing, including cache reads and writes. We compute these metrics separately at each switch point and average them within each reported window. Normalized metrics. Following Sec. 3.3, we evaluate each strategy at each switch point relative to the LC-only and HC-only baselines on its matched switched subset, placing all switch points on the same LC-to-HC reference scale. Let Rm(K)R_m(K) and Cm(K)C_m(K) denote pass rate and mean cost for strategy m; LC/HC subscripts denote the matched single-model baselines. Quality Recovery (QRec) is the fraction of HC’s quality advantage over LC that the strategy recovers, while Cost-Savings Retention (CSRet) is the fraction of LC’s cost advantage over HC that it retains: QRec⁡(m,K) (m,K) =100Rm(K)−RLC(K)RHC(K)−RLC(K), =100 R_m(K)-R_LC(K)R_HC(K)-R_LC(K), (1) CSRet⁡(m,K) (m,K) =100CHC(K)−Cm(K)CHC(K)−CLC(K). =100 C_HC(K)-C_m(K)C_HC(K)-C_LC(K). (2) QRec is 00 at LC quality and 100100 at HC quality; CSRet is 100100 at LC cost and 00 at HC cost. Values outside [0,100][0,100] fall beyond these anchors (e.g., negative CSRet is costlier than HC-only). Experimental scale. For each model family, we evaluate 58 configurations: two single-model baselines and 56 handoff configurations spanning seven switch points, four handoff strategies, and two directions. Running all configurations on the 500 SWE-bench Verified instances across both model families yields 58,000 agent runs, 2 million LLM API calls, and 36 billion processed tokens. 4 The Handoff Tax Claude (Haiku 4.5 / Opus 4.7) GPT-5.6 (Luna / Sol) Strategy Pass Cost Steps QRec CSRet Strategy Pass Cost Steps QRec CSRet (a) Escalation (LC → HC) LC-only 60.7 0.40 80 00 100100 LC-only 58.7 0.06 14 00 100100 HC-only 79.2 0.72 31 100100 00 HC-only 83.7 0.47 17 100100 00 Abort + HC fresh 79.2† 0.90 78 100†100 −58-58 Abort + HC fresh 83.7† 0.51 27 100†100 −11-11 LC-full + HC-full 79.2† 1.12 111 100†100 −130-130 LC-full + HC-full 83.7† 0.53 31 100†100 −15-15 Raw handoff 69.2 1.61 74 4747 −285-285 Raw handoff 67.5 0.36 17 3636 2626 Compactpre_pre 71.8 0.75 71 6060 −-11 Compactpre_pre 68.8 0.27 18 4040 49 Compactsuf_suf 69.6 0.98 67 4949 −82-82 Compactsuf_suf 68.8 0.43 19 4040 1010 Traj-drop 72.4 0.81 74 64 −30-30 Traj-drop 79.7 0.50 25 84 −8-8 (b) Downshift (HC → LC) LC-only 54.6 0.41 79 00 100100 LC-only 63.6 0.05 13 00 100100 HC-only 75.8 0.85 35 100100 00 HC-only 85.8 0.47 18 100100 00 Raw handoff 65.6 0.51 58 5050 80 Raw handoff 81.0 0.41 16 79 14 Compactpre_pre 66.8 0.52 54 56 7878 Compactpre_pre 79.8 0.43 18 7272 1010 Compactsuf_suf 63.7 0.53 65 4242 7373 Compactsuf_suf 80.4 0.42 17 7575 1313 Traj-drop 60.9 0.59 79 2828 5959 Traj-drop 75.4 0.43 24 5353 1010 Table 1: Aggregate coding-agent handoffs across model pairs. Raw escalation recovers less than half of HC’s quality advantage, whereas downshift offers a favorable intermediate cost–quality point. Reducing inherited LC context improves escalation, while preserving HC context improves downshift. Values are unweighted means over seven switch points on matched switched subsets; QRec and CSRet follow Sec. 3.4. Figure 3: Handoff effects vary with task difficulty. For Claude, all escalation interfaces are unfavorable on easy and medium tasks, but on hard tasks reduced-context interfaces become cheaper than HC-only while recovering 65–74% of HC’s quality advantage; Raw does not make this transition. In downshift, dropping the HC trajectory yields the worst cost–quality trade-off among tested interfaces in every difficulty bucket. We evaluate both switching directions under the protocol of Sec. 3 and report matched switched-subset results in Tab. 1. Across both model families, two qualitative patterns emerge. First, Raw handoff recovers less than half of HC’s quality advantage in escalation, but offers a favorable cost–quality trade-off in downshift at lower cost than HC-only. Second, inherited trajectory has direction-dependent value: dropping LC trajectory improves escalation, whereas preserving HC trajectory improves downshift. These quality patterns recur across families, while their cost implications differ because the model pairs have different baseline trajectory lengths and HC-to-LC per-task cost ratios. We examine escalation (Sec. 4.1), downshift (Sec. 4.2), and the receiver computations underlying these effects (Sec. 4.3). Sec. A.3 reports uncertainty estimates for the interface comparisons. 4.1 Escalation: Limited Rescue at High Cost Raw escalation fails to recover most of HC’s quality advantage. Raw is the default handoff interface, passing the full LC trajectory directly to HC. Across both model pairs in Panel (a) of Tab. 1, it recovers less than half of the LC–HC quality gap (QRec=47%=47\% for Claude and 36% for GPT). This limited quality recovery is common to both families; whether it also creates a severe cost premium depends on the model pair. Raw’s limited rescue comes at substantial cost. Raw escalation substantially increases cost over LC-only in both families, by approximately 4.0×4.0× for Claude and 6.1×6.1× for GPT. For Claude, this increase is large enough that Raw costs more than twice as much as starting with HC ($1.61 vs. $0.72). By contrast, Raw remains cheaper than HC-only for GPT ($0.36 vs. $0.47). This difference reflects the model pairs’ baseline per-task economics: HC-only costs roughly 8×8× as much as LC-only for GPT, compared with roughly 2×2× for Claude. For Claude, Raw’s limited rescue and severe cost premium motivate a stricter question: once LC work has been incurred, is it better to continue Raw or abandon the run and restart HC? We compare Raw with two controls that charge for LC’s work but discard both its conversation and working-tree edits before restarting HC from the original task state. Abort + HC fresh charges for LC through the switch point before running HC from scratch; LC-full + HC-full charges for a complete LC run before restarting HC. Both recover HC-only quality by construction. Even after charging for the discarded LC work, both controls remain cheaper than Raw ($0.90 and $1.12, respectively, versus $1.61). Thus, for the Claude pair, Raw escalation is strictly dominated by restarting HC from scratch: restarting costs less and solves more tasks. We next test whether reducing the inherited trajectory can improve Raw’s limited rescue without discarding LC’s working-tree edits. Different trajectory reductions favor different cost–quality objectives. Both compaction interfaces yield slightly higher observed pass rates than Raw, with differences of 0.4–2.6 points (Tab. 1). Compactpre provides the highest cost retention among the tested escalation interfaces in both families. For Claude, it raises QRec from 47% to 60% and improves CSRet from −285%-285\% to −11%-11\%, approaching HC-only cost parity. For GPT, QRec changes only slightly, from 36% to 40%, while CSRet improves from 26% to 49%. Taking context reduction further produces a larger quality gain. Traj-drop removes the trajectory while preserving LC’s working-tree edits. With Traj-drop, QRec rises from 47% to 64% for Claude and from 36% to 84% for GPT. This quality recovery comes with negative CSRet in both families (−30%-30\% and −8%-8\%), meaning that Traj-drop remains costlier than HC-only. Difficulty changes when Claude escalation is worthwhile. In aggregate, every Claude escalation interface costs more than HC-only, suggesting that escalation may never be worthwhile for this pair. We test whether this conclusion holds across task difficulty using the SWE-bench Verified labels. Fig. 3 reveals an exception on hard tasks. On easy tasks, all interfaces offer a poor cost–quality trade-off, recovering only around half of the LC–HC quality gap while costing far more than HC-only (CSRet −164%-164\% to −982%-982\%; Tab. 4). In the hard-task bucket, by contrast, all three reduced-context interfaces cost less than HC-only, retaining 12–42% of LC’s cost advantage while recovering 65–74% of the LC–HC quality gap. Raw does not make the same transition: it remains costlier than HC-only and recovers only 47% of the quality gap. Thus, difficulty alone does not make Raw escalation cost-effective, but it can make reduced-context escalation attractive. Because the hard subset is small (N¯≈24 N\!≈\!24 per cell), we treat this pattern as exploratory rather than as a cross-family finding. Takeaway. Raw escalation is a poor bargain: it recovers less than half of HC’s quality advantage at several times LC’s cost. For Claude, it even costs more than completing LC and then running HC from scratch. What HC inherits changes the trade-off: Compactpre minimizes handoff cost, while Traj-drop maximizes quality recovery. 4.2 Downshift: Context Preserves Quality Raw downshift lets LC build on HC’s work. In contrast to the limited rescue produced by Raw escalation, the same interface transfers useful HC progress when the direction reverses. Downshift tests whether LC can build on HC’s partial work. For Claude, Raw HC→ raises pass rate from 54.6% to 65.6%, while cost rises only from $0.41 to $0.51, retaining 80% of LC’s cost advantage. GPT shows a different balance: Raw retains 79% of HC’s quality advantage but only 14% of LC’s cost advantage. Thus, both LC receivers convert HC’s partial work into a favorable intermediate point below HC-only cost, with different balances: Claude retains most of LC’s cost advantage, whereas GPT retains most of HC’s quality advantage. Removing the HC trajectory harms downshift quality. Raw downshift transfers both HC’s working-tree edits and its trajectory, so its quality gain does not reveal which form of state helps LC. Traj-drop isolates the role of trajectory guidance by preserving the edits while removing the trajectory. For Claude, Traj-drop—the highest-quality escalation interface—becomes the lowest-quality and most expensive tested downshift strategy, recovering only 28% of the LC–HC quality gap, compared with 50% for Raw and 56% for Compactpre, while retaining just 59% of LC’s cost advantage. The difficulty breakdown in Fig. 3 shows that Traj-drop has the same quality-and-savings disadvantage across all difficulty levels. The same quality reversal appears for GPT. Traj-drop recovers only 53% of the LC–HC quality gap, compared with 79% for Raw and 72–75% for the two compaction interfaces. Cost retention spans only 10–14% across the GPT interfaces, so removing the trajectory primarily harms quality rather than savings. Thus, across both model families, the HC trajectory is important for downshift quality: preserving it retains substantially more of HC’s advantage than removing it entirely. Takeaway. Downshift offers a useful operating point: LC models retain a substantial share of HC’s quality advantage while remaining cheaper than HC-only. Removing the HC trajectory sharply reduces that retained quality. Together with escalation, this reveals a directional duality: dropping the sender’s trajectory helps an HC receiver but harms an LC receiver. 4.3 Cost Mechanics by Handoff Direction Having established that trajectory removal improves escalation quality but harms downshift quality across both families, we next examine how handoff direction changes receiver-side computation and cost. We decompose post-handoff cost into receiver step count and cost per step. The Claude and GPT decompositions are shown in Fig. 4 and Appendix A.3 (Fig. 6), respectively. Figure 4: The handoff tax arises through different computational mechanisms. In Claude escalation, Raw makes each post-handoff HC step 2.2×2.2× costlier than Compactpre; in downshift, Traj-drop makes LC take 1.6×1.6× more steps. Full LC context inflates HC calls, whereas missing HC context forces LC rework. Raw escalation inflates the cost of each HC step. Relative to Compactpre, Raw raises the average post-handoff cost per HC step by 2.2×2.2× for Claude and 1.6×1.6× for GPT, while the two interfaces require a similar number of HC steps within each family. The Raw-to-Compactpre escalation premium therefore arises primarily from more expensive receiver calls, not from a longer HC continuation. Traj-drop downshift takes additional LC steps. Relative to Compactpre, Traj-drop downshift requires 1.6×1.6× the post-handoff LC steps for Claude and 2.0×2.0× for GPT, while the two interfaces have similar costs per LC step within each family. The Traj-drop-to-Compactpre downshift premium therefore arises primarily from additional LC work, not from more expensive calls. This is consistent with reconstructing missing HC context. Takeaway. Raw escalation provides limited rescue, and the handoff tax depends on direction and interface. Removing inherited LC trajectory improves quality for HC receivers, whereas preserving inherited HC trajectory benefits LC receivers. The associated cost penalties manifest through different computational channels: more expensive HC steps under Raw escalation and additional LC steps under Traj-drop downshift. 5 Beyond the Coding-Agent Setting Our primary study focuses on coding agents, one of the most consequential real-world settings for long-running agents, using SWE-bench, where the specification is available upfront and execution leaves persistent repository state. We next vary an orthogonal property: the task’s information dynamics—how task-relevant information becomes available over time—which shape what the sender can accomplish before the switch and what remains for the receiver. As illustrated in Fig. 5, SWE-bench provides the specification upfront while repository information accumulates during execution; LiC Laban et al. (2025) reveals requirements incrementally across turns; and BrowseComp Wei et al. (2025) provides the question upfront while evidence accumulates through search. We evaluate Raw handoffs throughout to isolate this axis from interface design. Figure 5: Task-relevant information evolves differently across settings. SWE-bench provides the specification upfront, followed by rapid repository scoping; LiC reveals user requirements over turns; BrowseComp accumulates evidence through web search. The schematic illustrates what information is available at different stages of execution. Tasks Policy Score Cost QRec CSRet [-0.1em] LC-only 62.1 0.015 00 100100 HC-only 76.9 0.080 100100 00 Escalation 74.2 0.056 8686 3636 Downshift 67.0 0.046 3131 5353 Table 2: Late-arriving requirements favor an HC suffix. Raw-handoff results on LiC, averaged across five task families and early, middle, and late structural switch positions for Claude. Late-arriving requirements favor a strong suffix. We evaluate Raw handoffs on 535 LiC examples across five task families, where requirement shards progressively complete an initial high-level intent. Tab. 2 reports Claude averages across early, middle, and late handoffs; GPT and position-specific results, together with the full protocol, appear in Appendix B. Across all five families, late-arriving requirements reverse the coding quality ordering while preserving the cost ordering: escalation recovers 86% of HC’s quality advantage and retains 36% of LC’s cost advantage, versus 31% and 53% for downshift. By construction, LiC remains underspecified until the final requirement shard, so the receiver is the first model able to solve the fully specified task. This favors escalation, which assigns that decisive stage to HC. Strategy QRec CSRet Steps LC HC Full LC + full HC 100.0 −40.1-40.1 62.6 33.4 Abort + fresh HC 100.0 −8.5-8.5 23.0 33.4 Escalation 95.8 −30.0-30.0 23.0 30.4 Downshift 56.7 76.8 22.8 17.0 Table 3: Under progressive search, escalation recovers quality but not savings. GPT Raw-handoff results on BrowseComp, averaged across switch points. Progressive search recovers quality but not savings. We evaluate Raw handoffs on 200 BrowseComp questions filtered to require web browsing; the question is available upfront, but answer-relevant evidence accumulates through search. Tab. 3 reports GPT results. Full details in Appendix C. For GPT, escalation nearly closes the HC quality gap (QRec=95.8%QRec=95.8\%), unlike in SWE-bench, yet—as in SWE-bench—fails to produce savings: CSRet is −30.0%-30.0\%, worse than aborting LC and restarting HC from scratch (−8.5%-8.5\%). Downshift, by contrast, offers a useful middle ground, recovering 56.7% of HC’s quality advantage while retaining 76.8% of LC’s cost advantage. Escalation and Abort + fresh HC share the same 23-step LC prefix, but inheriting it shortens the HC continuation by three calls on average (30.4 vs. 33.4), suggesting that earlier search progress remains useful. Yet escalation remains substantially more expensive overall. Takeaway. Together, these extensions show that handoff value depends on information dynamics. LiC favors an HC receiver because the task becomes solvable only after late requirements arrive; BrowseComp shows that inherited search progress can nearly recover HC quality without producing savings. Thus, whether escalation is worthwhile depends not only on model capability, but also on the task-relevant state available at handoff. 6 Discussion and Conclusion Our results show that, under Raw handoff, escalation offers a poor cost–quality trade-off, whereas downshift yields a favorable intermediate cost–quality point. This directional asymmetry extends to interface design: reducing LC trajectory information improves escalation, whereas removing HC trajectory information harms downshift. Together, these suggest that model handoffs should be treated as a distinct inference problem rather than merely as an extension of model routing. Routing determines which model acts next, whereas the handoff interface determines what trajectory information that model inherits. In long-horizon coding agents, a handoff involves more than transferring conversational context: the receiver also inherits a repository state produced by another model. It must continue both an existing work product and a non-native trajectory, and how the trajectory is transferred alongside that persistent state can substantially affect both quality and cost. Thus, handoff design should be optimized jointly with receiver selection and switch timing. Future directions. We use fixed switch points to isolate the consequences of a handoff from the policy that triggers it; future work can study adaptive, progress-aware policies that jointly decide when to switch and which model should continue. Similarly, we evaluate a small set of controlled interfaces to separate full trajectory transfer, compaction, and trajectory removal; future work can develop structured handoffs that selectively preserve, summarize, or discard particular trajectory components. More broadly, routing methods should treat the handoff interface as part of the switching policy rather than assume that the full trajectory simply carries forward. Evaluation should also extend across model families, capability gaps, pricing regimes, repeated rollouts, and multiple handoffs per trajectory. Limitations We study two model pairs (Claude Haiku 4.5 / Opus 4.7 and GPT-5.6 Luna / Sol), and our primary coding-agent study uses one benchmark, SWE-bench Verified. Within this setting, however, we evaluate at substantial scale: 58 configurations per model family across all 500 tasks, totaling 58,000 agent runs, 2 million LLM API calls, and 36 billion processed tokens. To extend the scope of our study beyond the SWE-bench coding-agent setting, we evaluate a broader range of information dynamics but only under Raw transfer, so our interface findings are established primarily in the coding setting. Switch points are fixed in advance using model- and difficulty-calibrated percentiles of single-model trajectory length. The resulting matched switched subsets are relatively small for the hard SWE-bench stratum (N¯≈24 N\!≈\!24–2727 per reported cell), so we treat the difficulty-conditioned findings as exploratory. In the SWE-bench study, we run a single episode for each of the 500 tasks under every configuration and therefore do not estimate variability across repeated runs of the same task under the same configuration. To preserve comparability despite this limitation, we restrict each interface comparison to the shared intersection of tasks that switched under every strategy. We also report task-clustered bootstrap confidence intervals, preserving each task’s observations across switch points. Cost comparisons, difficulty analyses, and mechanism interpretations remain descriptive. Finally, dollar-cost conclusions depend on provider pricing and prompt-cache rates. References Aggarwal et al. (2025) P. Aggarwal, A. Madaan, A. Anand, S. P. Potharaju, S. Mishra, P. Zhou, A. Gupta, D. Rajagopal, K. Kappaganthu, Y. Yang, S. Upadhyay, M. Faruqui, and Mausam AutoMix: automatically mixing language models. External Links: 2310.12963, Link Cited by: §2. Anthropic (2025) Anthropic Claude haiku 4.5 system card. Note: https://w.anthropic.com/claude-haiku-4-5-system-cardAccessed: 2026-07-04 Cited by: §3. Anthropic (2026) Anthropic Claude opus 4.7 system card. Note: https://w.anthropic.com/claude-opus-4-7-system-cardAccessed: 2026-07-04 Cited by: §3. Chen et al. (2023) L. Chen, M. Zaharia, and J. Zou FrugalGPT: how to use large language models while reducing cost and improving performance. External Links: 2305.05176, Link Cited by: §2. Chen et al. (2021) M. Chen, J. Tworek, H. Jun, Q. Yuan, H. P. de Oliveira Pinto, J. Kaplan, H. Edwards, Y. Burda, N. Joseph, G. Brockman, A. Ray, R. Puri, G. Krueger, M. Petrov, H. Khlaaf, G. Sastry, P. Mishkin, B. Chan, S. Gray, N. Ryder, M. Pavlov, A. Power, L. Kaiser, M. Bavarian, C. Winter, P. Tillet, F. P. Such, D. Cummings, M. Plappert, F. Chantzis, E. Barnes, A. Herbert-Voss, W. H. Guss, A. Nichol, A. Paino, N. Tezak, J. Tang, I. Babuschkin, S. Balaji, S. Jain, W. Saunders, C. Hesse, A. N. Carr, J. Leike, J. Achiam, V. Misra, E. Morikawa, A. Radford, M. Knight, M. Brundage, M. Murati, K. Mayer, P. Welinder, B. McGrew, D. Amodei, S. McCandlish, I. Sutskever, and W. Zaremba Evaluating large language models trained on code. External Links: 2107.03374 Cited by: §B.1. Cobbe et al. (2021) K. Cobbe, V. Kosaraju, M. Bavarian, M. Chen, H. Jun, L. Kaiser, M. Plappert, J. Tworek, J. Hilton, R. Nakano, C. Hesse, and J. Schulman Training verifiers to solve math word problems. External Links: 2110.14168, Link Cited by: §B.1. Ding et al. (2024) D. Ding, A. Mallick, C. Wang, R. Sim, S. Mukherjee, V. Ruhle, L. V. S. Lakshmanan, and A. H. Awadallah Hybrid llm: cost-efficient and quality-aware query routing. External Links: 2404.14618, Link Cited by: §2. Hemadri et al. (2026) R. V. Hemadri, H. F. Mohammed, R. Maheshwary, S. Daruru, S. Davasam, V. Yadav, S. Sunkara, and S. Rajeswar R2V agent: teaching slms when to ask for help. External Links: 2605.16604, Link Cited by: §2. Hitzig et al. (2026) Z. Hitzig, M. Massenkoff, E. Lyubich, S. Zhang, R. Heller, and P. McCroryAgentic coding and persistent returns to expertise(Website) External Links: Link Cited by: §1. Hu et al. (2024) Q. J. Hu, J. Bieker, X. Li, N. Jiang, B. Keigwin, G. Ranganath, K. Keutzer, and S. K. Upadhyay RouterBench: a benchmark for multi-llm routing system. External Links: 2403.12031, Link Cited by: §2. Jain et al. (2024) N. Jain, K. Han, A. Gu, W. Li, F. Yan, T. Zhang, S. Wang, A. Solar-Lezama, K. Sen, and I. Stoica LiveCodeBench: holistic and contamination free evaluation of large language models for code. arXiv preprint arXiv:2403.07974. Cited by: §B.1. Jimenez et al. (2024) C. E. Jimenez, J. Yang, A. Wettig, S. Yao, K. Pei, O. Press, and K. R. Narasimhan SWE-bench: can language models resolve real-world github issues?. In The Twelfth International Conference on Learning Representations, External Links: Link Cited by: §1, §3. KC and Budathoki (2026) D. KC and A. Budathoki Handoff debt: the rediscovery cost when coding agents take over interrupted tasks. External Links: 2606.02875, Link Cited by: §2. Khraishi et al. (2026) R. Khraishi, I. Zafar, K. Myles, and G. A. Cowan Evaluating performance drift from model switching in multi-turn llm systems. External Links: 2603.03111, Link Cited by: §2. Kotte (2026) V. Kotte UCCI: calibrated uncertainty for cost-optimal llm cascade routing. External Links: 2605.18796, Link Cited by: §2. Laban et al. (2025) P. Laban, H. Hayashi, Y. Zhou, and J. Neville LLMs get lost in multi-turn conversation. External Links: 2505.06120, Link Cited by: §B.1, §5. Ong et al. (2025) I. Ong, A. Almahairi, V. Wu, W. Chiang, T. Wu, J. E. Gonzalez, M. W. Kadous, and I. Stoica RouteLLM: learning to route llms with preference data. External Links: 2406.18665, Link Cited by: §2. OpenAI (2026) OpenAI GPT-5.6: frontier intelligence that scales with your ambition. External Links: Link Cited by: §3. Parikh et al. (2020) A. Parikh, X. Wang, S. Gehrmann, M. Faruqui, B. Dhingra, D. Yang, and D. Das ToTTo: a controlled table-to-text generation dataset. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), B. Webber, T. Cohn, Y. He, and Y. Liu (Eds.), Online, p. 1173–1186. External Links: Link, Document Cited by: §B.1. Patil et al. (2023) S. G. Patil, T. Zhang, X. Wang, and J. E. Gonzalez Gorilla: large language model connected with massive apis. External Links: 2305.15334, Link Cited by: §1. Post (2018) M. Post A call for clarity in reporting BLEU scores. In Proceedings of the Third Conference on Machine Translation: Research Papers, Brussels, Belgium, p. 186–191. External Links: Link, Document Cited by: §B.1. Qin et al. (2023) Y. Qin, S. Liang, Y. Ye, K. Zhu, L. Yan, Y. Lu, Y. Lin, X. Cong, X. Tang, B. Qian, S. Zhao, L. Hong, R. Tian, R. Xie, J. Zhou, M. Gerstein, D. Li, Z. Liu, and M. Sun ToolLLM: facilitating large language models to master 16000+ real-world apis. External Links: 2307.16789, Link Cited by: §1. Schick et al. (2023) T. Schick, J. Dwivedi-Yu, R. Dessì, R. Raileanu, M. Lomeli, L. Zettlemoyer, N. Cancedda, and T. Scialom Toolformer: language models can teach themselves to use tools. External Links: 2302.04761, Link Cited by: §1. Son et al. (2026) S. Son, S. Yoon, J. Tang, S. Wang, L. Wolf, and I. Bogunovic SWE-router: routing in multi-turn agentic software engineering tasks. External Links: 2607.00053, Link Cited by: §2. Valkanas et al. (2025) A. Valkanas, S. Pal, P. Rumiantsev, Y. Zhang, and M. Coates C3PO: optimized large language model cascades with probabilistic cost constraints for reasoning. External Links: 2511.07396, Link Cited by: §2. Wang et al. (2024) X. Wang, Y. Chen, L. Yuan, Y. Zhang, Y. Li, H. Peng, and H. Ji Executable code actions elicit better LLM agents. In Proceedings of the 41st International Conference on Machine Learning, R. Salakhutdinov, Z. Kolter, K. Heller, A. Weller, N. Oliver, J. Scarlett, and F. Berkenkamp (Eds.), Proceedings of Machine Learning Research, Vol. 235, p. 50208–50232. External Links: Link Cited by: §1. Wang et al. (2025) 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 OpenHands: an open platform for ai software developers as generalist agents. External Links: 2407.16741, Link Cited by: §1. Wei et al. (2025) J. Wei, Z. Sun, S. Papay, S. McKinney, J. Han, I. Fulford, H. W. Chung, A. T. Passos, W. Fedus, and A. Glaese BrowseComp: a simple yet challenging benchmark for browsing agents. External Links: 2504.12516, Link Cited by: Appendix C, §5. Xu et al. (2025) F. F. Xu, Y. Song, B. Li, Y. Tang, K. Jain, M. Bao, Z. Z. Wang, X. Zhou, Z. Guo, M. Cao, M. Yang, H. Y. Lu, A. Martin, Z. Su, L. Maben, R. Mehta, W. Chi, L. Jang, Y. Xie, S. Zhou, and G. Neubig TheAgentCompany: benchmarking llm agents on consequential real world tasks. External Links: 2412.14161, Link Cited by: §1. Yan et al. (2024) F. Yan, H. Mao, C. C. Ji, T. Zhang, S. G. Patil, I. Stoica, and J. E. Gonzalez Berkeley function calling leaderboard. Note: https://gorilla.cs.berkeley.edu/leaderboard.html Cited by: §B.1. Yang et al. (2024) J. Yang, C. E. Jimenez, A. Wettig, K. Lieret, S. Yao, K. Narasimhan, and O. Press SWE-agent: agent-computer interfaces enable automated software engineering. External Links: 2405.15793, Link Cited by: §3. Yao et al. (2023) S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. Narasimhan, and Y. Cao ReAct: synergizing reasoning and acting in language models. External Links: 2210.03629, Link Cited by: §1. Yu et al. (2018) T. Yu, R. Zhang, K. Yang, M. Yasunaga, D. Wang, Z. Li, J. Ma, I. Li, Q. Yao, S. Roman, Z. Zhang, and D. Radev Spider: a large-scale human-labeled dataset for complex and cross-domain semantic parsing and text-to-SQL task. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, E. Riloff, D. Chiang, J. Hockenmaier, and J. Tsujii (Eds.), Brussels, Belgium, p. 3911–3921. External Links: Link, Document Cited by: §B.1. Yue et al. (2024) M. Yue, J. Zhao, M. Zhang, L. Du, and Z. Yao Large language model cascades with mixture of thoughts representations for cost-efficient reasoning. External Links: 2310.03094, Link Cited by: §2. Zhang et al. (2026a) J. Zhang, X. Liu, Y. Hu, C. Niu, H. Zeng, S. Tang, F. Wu, and G. Chen From myopic selection to long-horizon awareness: sequential llm routing for multi-turn dialogue. External Links: 2604.12385, Link Cited by: §2. Zhang et al. (2026b) Y. Zhang, H. Li, Z. Wang, S. Feng, X. Yang, D. Wang, B. Zhang, L. Bai, and S. Hu MTRouter: cost-aware multi-turn llm routing with history-model joint embeddings. External Links: 2604.23530, Link Cited by: §2. Zhong et al. (2020) R. Zhong, T. Yu, and D. Klein Semantic evaluation for text-to-SQL with distilled test suites. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), Online, p. 396–411. External Links: Link, Document Cited by: §B.1. Appendix A SWE-bench Verified A.1 Experimental Details We provide the full SWE-bench Verified protocol, including the agent environment, handoff implementation and prompts, switch-point calibration, cost accounting, and software configuration. Agent and task environment. We implement handoff by extending mini-SWE-agent with a wrapper containing exactly two models: a prefix model MpreM_pre and a suffix model MsufM_suf. A single agent loop and Docker environment are retained throughout the run. Consequently, the repository state, including all edits made by MpreM_pre, persists after handoff; the interfaces differ only in the trajectory information supplied to MsufM_suf. All models receive the same SWE-bench task prompt and the same Bash tool, and command outputs are returned to the model as tool observations. We evaluate all 500 test instances of SWE-bench Verified. Models and handoff directions. For Claude, the low-cost (LC) and high-capability (HC) models are Claude Haiku 4.5 and Claude Opus 4.7, respectively. For GPT, they are GPT-5.6 Luna and GPT-5.6 Sol. We study both escalation, MLC→MHCM_LC\!→ M_HC, and downshift, MHC→MLCM_HC\!→ M_LC. Claude models were accessed through LiteLLM’s chat-completions interface, whereas GPT models were accessed through Bedrock’s OpenAI-compatible Responses endpoint; both used the same Bash tool, with conversation histories translated into each endpoint’s native tool-call format. For sampling, we set temperature to 0 for Haiku; for GPT-5.6 and Opus 4.7, we use provider defaults, with medium and high reasoning effort, respectively. Switch rule. Let q=0,1,…q=0,1,… index agent model calls and let sbs_b be the switch point for difficulty bucket b. Calls q<sbq<s_b use MpreM_pre, and call q=sbq=s_b is the first call to MsufM_suf. Thus, sbs_b prefix actions and their resulting observations are complete before the suffix model is queried. Thresholds are fixed before the handoff runs at percentiles p∈5,10,15,25,35,45,50p∈\5,10,15,25,35,45,50\ of the prefix model’s single-model termination-step distribution. They are computed separately for easy (<15<15 minute), medium (1515 minutes–11 hour), and hard (>1>1 hour) tasks; the two longest SWE-bench difficulty categories are merged into the hard bucket. Tab. 11 report the exact thresholds. Handoff interfaces. Let 1:KT_1:K denote the trajectory produced by the prefix model up to step K, and let KW_K denote the persistent working tree at that point. All interfaces preserve KW_K; they differ only in the trajectory information passed to the suffix model. We compare four principal interfaces: Raw. The suffix model receives the full trajectory 1:KT_1:K together with KW_K: the system prompt, task, prefix reasoning, tool calls, and tool observations are retained verbatim. No message announces the model change. Compactpre. The prefix model reads 1:KT_1:K and writes a plain-text continuation summary. The suffix model receives only this summary together with KW_K. Compactsuf. The suffix model reads 1:KT_1:K and writes the continuation summary itself, allowing us to isolate which model should author the handoff representation. It then continues from this summary together with KW_K. Traj-drop. No trajectory information from 1:KT_1:K is transferred, and no summarization call is made. The suffix model receives only the system prompt, original task, a static continuation message, and KW_K. For both compact interfaces, the summarizer is given the full prefix trajectory followed by: Compaction instruction You are summarizing a coding agent trajectory for handing off work to another coding agent. The conversation above is a coding agent’s work-in-progress trajectory on a task. Write a summary of what the agent did so a different agent---with no access to this conversation---can continue the task. Let ^1:K T_1:K denote the resulting summary. We then replace the live context with [msystem,mtask,mhandoff(^1:K)],[\,m_system,\,m_task,\,m_handoff( T_1:K)\,], where mhandoffm_handoff uses the following message: Compaction continuation message Below is a summary of a previous agent’s work on this task. Any file changes it made are still in the working tree. <<SUMMARY>> Here, <<SUMMARY>> is replaced by ^1:K T_1:K. The suffix model then produces its first executable action. Traj-drop uses the same three-message structure but replaces the summary with the following fixed continuation message: Traj-drop continuation message A previous agent worked on this task. Any file changes it made are still in the working tree. Continue solving the task. The removed trajectory is written to an audit artifact but is never exposed to the suffix model. The restart controls are composed from existing matched runs rather than executed as separate trajectories. Abort + HC fresh combines the observed LC prefix cost and steps through K with the outcome, cost, and steps of the matched HC-only run. LC-full + HC-full similarly combines the cost and steps of the complete LC-only and HC-only runs. In both controls, quality is given by the matched HC-only outcome. Accounting. Each response is tagged with its active model, zero-indexed call number, prefix/suffix phase, effective threshold, and task difficulty. The compact interfaces make one paid summarization call; its cost and usage are attributed to the handoff and folded into the first suffix step, although it is not counted as an additional environment-action step. Traj-drop and Raw introduce no additional model calls. Costs are reconstructed from provider-reported input, cache, and output-token usage using the corresponding cache-aware price cards. Output-token counts include reasoning tokens. Billable malformed responses are retained and counted. Runs use a 150-step cap and are audited so that prefix, handoff, and suffix costs reconcile exactly with total instance cost. Final patches are scored with the official SWE-bench evaluator. Pricing and audit. Token prices follow LiteLLM v1.83.14’s model-pricing registry for the configured Bedrock model identifiers, including cache pricing and cross-region premiums where applicable. The GPT-5.6 Bedrock identifiers use a pinned LiteLLM-compatible local price card distributed with our experiment configuration. The accompanying artifact records the exact invocation identifiers, price-card entries, and implementation and evaluator revisions. Before reporting results, we verify that every required task–condition run has an audited terminal trajectory and evaluator outcome. Infrastructure failures are rerun rather than scored, and no task or condition is excluded. Aggregate retry-attempt counts were not retained. A.2 Switch-Point Calibration Tab. 11 reports the exact difficulty-calibrated switch steps used in the SWE-bench experiments for both model families and handoff directions. Each threshold is the corresponding percentile of the prefix model’s single-model termination-step distribution within a difficulty bucket. Escalation therefore uses thresholds calibrated from LC trajectories, whereas downshift uses thresholds calibrated from HC trajectories. Figure 6: GPT exhibits the same handoff-cost accounting pattern as Claude. Relative to Compactpre, each post-handoff HC step under Raw costs 1.6×1.6× as much, while Traj-drop downshift takes 2.0×2.0× as many post-handoff LC steps. A.3 Additional Results and Analyses Full results across model families. Tab. 1 reports the complete aggregate results for the Claude and GPT model pairs on matched switched subsets. Across both families, Raw escalation recovers less than half of the LC–HC quality gap, whereas downshift occupies a favorable intermediate cost–quality point. Reducing LC trajectory information improves escalation quality, while preserving HC trajectory information improves downshift quality. The analyses below disaggregate and test these aggregate patterns. Difficulty breakdown. Tab. 4 conditions the Claude results on the SWE-bench Verified difficulty buckets, providing the numbers behind Fig. 3. Escalation economics improve with difficulty, with the structured interfaces recovering more quality at a more favorable cost. In downshift, the context-preserving interfaces consistently outperform Traj-drop in quality recovery and cost efficiency. Results for the small hard-task subset should be interpreted directionally. (a) Escalation Strategy Difficulty N¯ N Pass ΔLC _LC ΔHC _HC QRec CSRet (%) (p) (p) (%) (%) Raw Easy 129 82.1 +3.9 -6.0 38 -982 Medium 160 63.6 +10.8 -11.7 49 -275 Hard 24 32.5 +19.2 -21.4 47 -86 Compactpre_pre Easy 129 83.4 +5.2 -4.7 53 -164 Medium 160 66.1 +13.3 -9.2 59 -12 Hard 24 42.7 +29.5 -11.2 74 42 Compactsuf_suf Easy 129 82.2 +4.1 -5.8 41 -482 Medium 160 63.1 +10.3 -12.2 47 -68 Hard 24 39.6 +26.4 -14.3 65 12 Traj-drop Easy 129 81.9 +3.7 -6.2 38 -213 Medium 160 68.6 +15.9 -6.7 71 -31 Hard 24 42.3 +29.0 -11.6 72 34 (b) Downshift Strategy Difficulty N¯ N Pass ΔLC _LC ΔHC _HC QRec CSRet (%) (p) (p) (%) (%) Raw Easy 113 78.0 +3.0 -7.6 23 46 Medium 157 61.8 +15.2 -9.8 59 81 Hard 27 35.6 +20.6 -22.9 46 88 Compactpre_pre Easy 113 79.0 +4.0 -6.6 35 44 Medium 157 62.5 +15.8 -9.2 62 80 Hard 27 40.4 +25.4 -18.1 57 87 Compactsuf_suf Easy 113 76.2 +1.2 -9.4 6 28 Medium 157 59.1 +12.4 -12.5 48 74 Hard 27 38.8 +23.7 -19.8 56 86 Traj-drop Easy 113 72.8 -2.2 -12.8 -28 -31 Medium 157 57.4 +10.7 -14.2 42 59 Hard 27 31.3 +16.2 -27.2 38 82 Table 4: Difficulty-conditioned handoff results. Claude intersection-only results averaged uniformly over switch points p5–p50, split by SWE-bench Verified difficulty. ΔLC _LC is the pass-rate gain over LC-only; ΔHC _HC is the pass-rate difference from HC-only, with negative values indicating remaining gap. QRec and CSRet follow Eqs. 1–2. Bold marks the best ΔLC _LC, QRec, and CSRet within each difficulty bucket and direction. Early and late switch breakdown. Tabs. 5 and 6 split the Claude and GPT results into Early (p5–p15) and Late (p25–p50) switch windows. The central interface reversal is stable across both timing windows and model families: Traj-drop achieves the highest quality recovery in escalation and the lowest in downshift. Switch timing changes the magnitude, but not the direction, of this effect. Later escalation generally recovers less quality, whereas later downshift improves quality recovery for Raw and both compaction interfaces while retaining less savings. The finer ordering among the context-preserving interfaces varies, particularly for GPT. Claude (Haiku 4.5 / Opus 4.7) GPT-5.6 (Luna / Sol) Strategy Pass Cost Steps QRec CSRet Strategy Pass Cost Steps QRec CSRet All LC-only 60.7 0.40 80 00 100100 LC-only 58.7 0.06 14 00 100100 HC-only 79.2 0.72 31 100100 00 HC-only 83.7 0.47 17 100100 00 Abort + HC fresh 79.2† 0.90 78 100†100 −58-58 Abort + HC fresh 83.7† 0.51 27 100†100 −11-11 LC-full + HC-full 79.2† 1.12 111 100†100 −130-130 LC-full + HC-full 83.7† 0.53 31 100†100 −14-14 Raw handoff 69.2 1.61 74 4747 −285-285 Raw handoff 67.5 0.36 17 3636 2727 Compactpre_pre 71.8 0.75 71 6060 −-11 Compactpre_pre 68.8 0.27 18 4040 49 Compactsuf_suf 69.6 0.98 67 4949 −82-82 Compactsuf_suf 68.8 0.43 19 4040 1010 Traj-drop 72.4 0.81 74 64 −30-30 Traj-drop 79.7 0.50 25 84 −8-8 Early LC-only 63.5 0.36 73 00 100100 LC-only 63.2 0.05 13 00 100100 HC-only 80.4 0.64 29 100100 00 HC-only 83.7 0.42 17 100100 00 Abort + HC fresh 80.4† 0.78 65 100†100 −49-49 Abort + HC fresh 83.7† 0.46 25 100†100 −10-10 LC-full + HC-full 80.4† 1.00 102 100†100 −129-129 LC-full + HC-full 83.7† 0.47 29 100†100 −13-13 Raw handoff 73.1 1.38 64 5757 −265-265 Raw handoff 72.3 0.32 16 4444 2727 Compactpre_pre 74.7 0.66 60 6666 −-6 Compactpre_pre 71.7 0.24 16 4242 49 Compactsuf_suf 73.6 0.82 56 6060 −65-65 Compactsuf_suf 71.4 0.37 17 4040 1515 Traj-drop 75.0 0.73 62 69 −32-32 Traj-drop 81.2 0.45 23 87 −8-8 Late LC-only 58.6 0.44 84 00 100100 LC-only 55.4 0.07 15 00 100100 HC-only 78.3 0.77 34 100100 00 HC-only 83.8 0.50 18 100100 00 Abort + HC fresh 78.3† 0.99 87 100†100 −66-66 Abort + HC fresh 83.8† 0.55 28 100†100 −12-12 LC-full + HC-full 78.3† 1.21 118 100†100 −131-131 LC-full + HC-full 83.8† 0.56 32 100†100 −15-15 Raw handoff 66.4 1.78 82 3939 −299-299 Raw handoff 63.9 0.39 18 3030 2626 Compactpre_pre 69.6 0.82 79 5656 −-16 Compactpre_pre 66.6 0.29 18 3939 49 Compactsuf_suf 66.6 1.09 74 4040 −95-95 Compactsuf_suf 66.9 0.47 20 4141 77 Traj-drop 70.4 0.87 82 60 −28-28 Traj-drop 78.6 0.53 27 82 −8-8 Table 5: Coding-agent escalation by switch-timing window. Columns compare model families. All averages p5–p50, Early averages p5,p10,p15, and Late averages p25,p35,p45,p50. Values are unweighted means over switch points on matched switched subsets. Pass is in percent, Cost in dollars, and Steps is the mean trajectory length. QRec is recovered LC–HC quality gap; CSRet is retained cost savings, with negative values costlier than HC-only. Gray rows are single-model baselines and restart references. Bold marks the best available handoff quality and best non-baseline cost/efficiency within each family and timing window. Claude (Haiku 4.5 / Opus 4.7) GPT-5.6 (Luna / Sol) Strategy Pass Cost Steps QRec CSRet Strategy Pass Cost Steps QRec CSRet All LC-only 54.6 0.41 79 00 100100 LC-only 63.6 0.05 13 00 100100 HC-only 75.8 0.85 35 100100 00 HC-only 85.8 0.47 18 100100 00 Raw handoff 65.6 0.51 58 5050 80 Raw handoff 81.0 0.42 16 79 14 Compactpre_pre 66.8 0.52 54 56 7878 Compactpre_pre 79.8 0.43 18 7272 1010 Compactsuf_suf 63.7 0.53 65 4242 7373 Compactsuf_suf 80.4 0.42 17 7575 1313 Traj-drop 60.9 0.59 79 2828 5959 Traj-drop 75.4 0.43 24 5353 1010 Early LC-only 62.5 0.36 72 00 100100 LC-only 66.5 0.05 12 00 100100 HC-only 80.3 0.66 30 100100 00 HC-only 86.0 0.41 17 100100 00 Raw handoff 69.0 0.41 55 3535 83 Raw handoff 82.3 0.34 15 81 2020 Compactpre_pre 71.0 0.41 49 47 83 Compactpre_pre 79.9 0.36 17 6969 1616 Compactsuf_suf 67.6 0.44 60 2828 7575 Compactsuf_suf 79.9 0.34 16 6969 21 Traj-drop 65.9 0.48 72 1818 5959 Traj-drop 77.2 0.35 22 5555 1818 Late LC-only 48.6 0.45 85 00 100100 LC-only 61.4 0.06 14 00 100100 HC-only 72.4 0.98 40 100100 00 HC-only 85.7 0.51 19 100100 00 Raw handoff 63.1 0.58 60 6161 77 Raw handoff 79.9 0.47 17 7777 9 Compactpre_pre 63.6 0.59 58 63 7474 Compactpre_pre 79.8 0.49 19 7575 55 Compactsuf_suf 60.9 0.60 69 5252 7171 Compactsuf_suf 80.7 0.48 18 80 88 Traj-drop 57.1 0.68 85 3636 5959 Traj-drop 74.0 0.50 26 5252 44 Table 6: Coding-agent downshift by switch-timing window. Columns compare model families. All averages p5–p50, Early averages p5,p10,p15, and Late averages p25,p35,p45,p50. Values are unweighted means over switch points on matched switched subsets. Pass is in percent, Cost in dollars, and Steps is the mean trajectory length. QRec is recovered LC–HC quality gap; CSRet is retained cost savings. Gray rows are single-model baselines. Bold marks the best available handoff quality and best non-baseline cost/efficiency within each family and timing window. Receiver-side computation. Figure 6 reports the GPT decomposition corresponding to the Claude analysis in Sec. 4.3. The same pattern holds across model families. In escalation, Raw increases the average cost of each post-handoff HC step relative to Compactpre, while requiring a similar number of HC steps. In downshift, Traj-drop requires substantially more post-handoff LC steps than Compactpre, while the interfaces have similar costs per LC step. Thus, the Raw escalation premium arises primarily from more expensive receiver calls, whereas the Traj-drop downshift premium arises primarily from additional LC work. Model-change disclosure ablation. In raw handoff, a natural question is whether simply telling the receiver that the preceding work was produced by a different model recovers part of the escalation tax. To isolate the effect of model-change disclosure from context transformation, we compare Raw escalation with a disclosure-only variant using Claude. Both conditions preserve the full trajectory without summarization; the disclosure condition appends a single user message immediately before the suffix model’s first call: Handoff disclosure Injected user message The preceding assistant messages contain a previous agent’s work on this task. The previous agent used a different model. Any file changes it made are still in the working tree. Continue solving the original task. Across p5–p50, averaged uniformly over switch points on the shared fired intersection—here recomputed to include the disclosure condition, so the baselines differ slightly from Tab. 1—disclosure provides a modest quality gain at a slightly higher cost (Tab. 7). It remains below HC-only and is dominated in both quality and cost by Compactpre_pre and Traj-drop. This disclosure-only prompt does not eliminate the Raw handoff tax in Claude escalation. Strategy Pass Cost Steps QRec CSRet (%) ($) (#) (%) (%) LC-only 60.6 0.41 80 00 100100 HC-only 78.8 0.72 31 100100 00 Raw handoff 69.3 1.62 74 4848 −285-285 Raw + disclosure 70.5 1.67 75 5555 −300-300 Compactpre_pre 71.5 0.75 71 6060 −-10 Compactsuf_suf 69.6 0.98 67 5050 −83-83 Traj-drop 72.1 0.94 77 64 −72-72 Table 7: Escalation (LC → HC), averaged over all switch points. Intersection-only results average p5–p50 unweighted. QRec is quality-gap recovery and CSRet is cost-savings retention; both are 100% at the ideal corner. LC=Haiku 4.5 and HC=Opus 4.7. Compactpre_pre/Compactsuf_suf use the handoff-from/handoff-to model to summarize. Raw + disclosure retains the full trajectory and appends one model-change notice at the first HC call. Best Pass/QRec among handoff strategies and best Cost/Steps/CSRet among all non-baseline methods are in bold. Statistical uncertainty. We quantify task-to-task variation in pairwise interface pass-rate differences. At each switch point, differences are computed on the shared subset of tasks that reached the handoff under all four interfaces. The seven switch-point differences are then averaged with equal weight, matching the aggregation in the main table. For each model family and direction, we collect the unique task IDs appearing in at least one of the seven matched subsets. These pools contain 460 escalation and 458 downshift tasks for Claude, and 411 escalation and 426 downshift tasks for GPT. Each bootstrap replicate draws the same number of task IDs with replacement. When a task is drawn more than once, it receives the same weight under all interfaces and at every switch point where it appears. We then recompute the seven pass-rate differences and their equal-weight average. We report pointwise 95% confidence intervals from 10,00010,000 bootstrap replicates. Tab. 8 presents the central Traj-drop–Raw contrasts first within each direction, followed by Traj-drop–compaction contrasts and secondary comparisons among the context-preserving interfaces. The central reversal is consistent across both families: Traj-drop improves escalation pass rate and reduces downshift pass rate relative to Raw, with all four pointwise intervals excluding zero. The broader Traj-drop–compaction contrasts follow the same directional pattern, whereas comparisons among Raw and the two compaction interfaces vary across families. In particular, we do not treat summary authorship as a cross-family finding. Direction Comparison Claude GPT First Second Δ 95% CI Δ 95% CI Escalation Traj-drop Raw +3.1+3.1 [+1.3,+5.0][+1.3,+5.0] +12.2+12.2 [+9.2,+15.4][+9.2,+15.4] Traj-drop Compactpre_pre +0.6+0.6 [−1.1,+2.3][-1.1,+2.3] +10.9+10.9 [+7.5,+14.5][+7.5,+14.5] Traj-drop Compactsuf_suf +2.8+2.8 [+1.0,+4.6][+1.0,+4.6] +10.9+10.9 [+7.7,+14.2][+7.7,+14.2] Compactpre_pre Raw +2.5+2.5 [+0.9,+4.2][+0.9,+4.2] +1.3+1.3 [−1.7,+4.3][-1.7,+4.3] Compactsuf_suf Raw +0.4+0.4 [−1.2,+1.9][-1.2,+1.9] +1.3+1.3 [−1.4,+4.0][-1.4,+4.0] Compactpre_pre Compactsuf_suf +2.2+2.2 [+0.7,+3.7][+0.7,+3.7] −0.0-0.0 [−2.4,+2.4][-2.4,+2.4] Downshift Traj-drop Raw −4.7-4.7 [−7.3,−2.1][-7.3,-2.1] −5.6-5.6 [−8.6,−2.6][-8.6,-2.6] Traj-drop Compactpre_pre −5.9-5.9 [−8.5,−3.4][-8.5,-3.4] −4.4-4.4 [−7.1,−1.8][-7.1,-1.8] Traj-drop Compactsuf_suf −2.9-2.9 [−5.3,−0.4][-5.3,-0.4] −5.0-5.0 [−7.8,−2.3][-7.8,-2.3] Compactpre_pre Raw +1.2+1.2 [−0.6,+3.0][-0.6,+3.0] −1.1-1.1 [−3.2,+1.0][-3.2,+1.0] Compactsuf_suf Raw −1.9-1.9 [−4.0,+0.3][-4.0,+0.3] −0.6-0.6 [−2.7,+1.6][-2.7,+1.6] Compactpre_pre Compactsuf_suf +3.0+3.0 [+1.1,+5.1][+1.1,+5.1] −0.6-0.6 [−2.4,+1.3][-2.4,+1.3] Table 8: Pairwise interface pass-rate differences by model family. Δ is the first strategy minus the second, in percentage points. Within each direction, rows present the central Traj-drop–Raw reversal, broader Traj-drop–compaction comparisons, and secondary comparisons among context-preserving interfaces, in that order. Intervals are pointwise 95% task-clustered bootstrap confidence intervals from 10,00010,000 resamples. Appendix B Lost in Conversation B.1 Dataset and Protocol Corpus and task families. We build on the sharded-instruction corpora of Lost in Conversation (LiC) (Laban et al., 2025), which decompose each fully specified task into an ordered list of requirement shards, revealed one per user turn. We use five task families: Code (45 HumanEval (Chen et al., 2021) and 55 LiveCodeBench (Jain et al., 2024); n=100n=100), Database (Spider (Yu et al., 2018); n=107n=107), Actions (the parallel category of the Berkeley Function-Calling Leaderboard (Yan et al., 2024); n=105n=105), Math (GSM8K (Cobbe et al., 2021); n=103n=103), and Data-to-text (ToTTo (Parikh et al., 2020); n=120n=120), for 535 tasks total. We exclude the released summarization family because its official evaluator is an external LLM judge. Tasks contain N∈[3,12]N∈[3,12] shards, with per-family medians of 4–7. Deterministic sharded protocol. The original LiC protocol uses an LLM user simulator to paraphrase and schedule shards. We instead reveal one released shard per user turn, verbatim and in the official order, with one assistant call after every shard and no early termination. The final user turn appends a fixed family-specific instruction requesting the official answer format; only the final assistant response is evaluated. Every episode therefore contains exactly N assistant turns, and its user-turn sequence is byte-identical across models, policies, and switch points. Models and conversations. We evaluate the same two within-family pairs as in the coding study: Claude Haiku 4.5/Opus 4.7 and GPT-5.6 Luna/Sol, with the same model identifiers and inference settings as in the coding study. LiC is chat-only and uses no tools. Assistant turns are re-fed as plain visible text without hidden reasoning blocks, making the inherited transcript model-agnostic. Each task–policy cell contains one episode; the schedule and evaluators are deterministic, while model generation is not. Raw handoff and switch timing. At switch point K, the sender answers shards 1,…,K1,…,K, and the receiver answers shards K+1,…,NK+1,…,N in the same conversation. The receiver inherits the full visible history, including the sender’s responses, and no message announces the model change. Because episode length is fixed, switches fire on every task. We use Early, K=1K=1; Middle, K=⌊N/2⌋K= N/2 ; and Late, K=N−1K=N-1. Evaluation and metrics. We use the official LiC evaluators: numeric exact match for Math, the official AST checker for Actions, execution match against the Spider test-suite databases (Zhong et al., 2020) for Database, sandboxed reference tests for Code, and multi-reference sacreBLEU (Post, 2018) for Data-to-text. “Score” is pass rate times 100 except for Data-to-text, where it is mean per-task BLEU times 100. QRec and CSRet use matched single-model anchors under the same sharded schedule. Costs sum sender-priced prefix and receiver-priced suffix calls, including cache reads and writes. In disaggregated results, QRec is omitted whenever the absolute LC–HC anchor gap is below five points, where the normalized denominator is unstable. This rule omits Claude Math and GPT Database and Math. Aggregate QRec includes all five task families. B.2 Results by Switch Position Tab. 9 disaggregates the Claude results by structural switch position. Later escalation retains more savings, while quality is mostly stable. Later downshift generally improves quality while retaining less savings. Tab. 10 reports the GPT replication. The timing–direction interaction is similar: later escalation retains more savings, whereas later downshift retains less. The main departure is Actions downshift, where the small LC–HC anchor gap makes QRec volatile and the handoff falls below LC-only at the middle and late positions. Appendix C BrowseComp Benchmark and task selection. We use BrowseComp (Wei et al., 2025), a benchmark of 1,266 difficult information-seeking questions with exact reference answers. To focus on tasks that require browsing, we first sample 300 questions stratified by topic. We remove 22 questions that a prior Claude Haiku–Opus screen answered successfully without browsing, then select a fixed cohort of 200 questions, again stratified by topic, from the remaining 278. Because the screening used the Claude pair, we do not claim that the resulting cohort is specifically decontaminated for the GPT-5.6 models evaluated here. Models and browsing harness. As in the coding study, we use GPT-5.6 Luna and GPT-5.6 Sol as LC and HC, respectively. Each model operates in a single-context, ReAct-style loop with web-search and URL-retrieval tools, where one assistant inference counts as one step. Episodes allow up to 100 steps, followed when needed by a browsing-disabled finalization call. Research stops at approximately 80% context utilization, preserving capacity for pending tool output and finalization. Raw handoff and switch timing. At switch step K, the receiver inherits the sender’s complete response prefix, including reasoning items, tool calls, and tool outputs. Prefix tools are not re-executed, no message announces the change, and both models share the original step budget. Switches fire only if the sender’s recorded trajectory continues beyond K. As in the SWE setup, we calibrate sender-specific switch points using percentiles. The reported p25 and p50 points are K=15,31K=15,31 for escalation and K=10,24K=10,24 for downshift. Each result uses the switched-only subset. Evaluation and metrics. Answers are graded with the official BrowseComp judge prompt and parsing rule, using Claude Opus 4.7 as judge; grader costs are excluded. QRec and CSRet use matched LC-only and HC-only anchors separately at each switch point and are then averaged unweighted over p25 and p50. Full LC + full HC pays for both complete trajectories and uses the HC outcome. Abort + fresh HC pays for the LC prefix followed by a fresh complete HC trajectory. Run accounting. Infrastructure failures are retried with backoff; exhausted episodes are rerun from scratch and never scored partially. Each task–policy cell contains one successfully audited generation, so we interpret these results descriptively. Retry attempts were not retained as a separate aggregate count. Task Policy Early Middle Late Score Cost QRec CSRet Score Cost QRec CSRet Score Cost QRec CSRet ($) (%) (%) ($) (%) (%) ($) (%) (%) Code LC-only 65.0 0.033 00 100100 65.0 0.033 00 100100 65.0 0.033 00 100100 HC-only 92.0 0.163 100100 00 92.0 0.163 100100 00 92.0 0.163 100100 00 Escalation 88.0 0.145 8585 1414 89.0 0.133 8989 2323 80.0 0.065 5656 7575 Downshift 75.0 0.047 3737 8989 85.0 0.082 7474 6262 85.0 0.156 7474 55 Database LC-only 51.4 0.011 00 100100 51.4 0.011 00 100100 51.4 0.011 00 100100 HC-only 72.0 0.061 100100 00 72.0 0.061 100100 00 72.0 0.061 100100 00 Escalation 67.3 0.068 7777 −12-12 65.4 0.049 6868 2525 67.3 0.024 7777 7373 Downshift 50.5 0.019 −5-5 8585 50.5 0.032 −5-5 5858 55.1 0.050 1818 2222 Actions LC-only 57.1 0.006 00 100100 57.1 0.006 00 100100 57.1 0.006 00 100100 HC-only 70.5 0.035 100100 00 70.5 0.035 100100 00 70.5 0.035 100100 00 Escalation 73.3 0.032 121121 1010 69.5 0.024 9393 3838 70.5 0.013 100100 7777 Downshift 59.0 0.010 1414 8585 61.0 0.018 2929 5858 61.0 0.029 2929 2222 Math LC-only 90.3 0.008 – 100100 90.3 0.008 – 100100 90.3 0.008 – 100100 HC-only 94.2 0.052 – 00 94.2 0.052 – 00 94.2 0.052 – 00 Escalation 94.2 0.049 – 66 94.2 0.040 – 2727 94.2 0.018 – 7777 Downshift 89.3 0.013 – 9090 91.3 0.023 – 6666 94.2 0.043 – 2020 Data-to-text LC-only 46.5 0.015 00 100100 46.5 0.015 00 100100 46.5 0.015 00 100100 HC-only 55.6 0.089 100100 00 55.6 0.089 100100 00 55.6 0.089 100100 00 Escalation 55.4 0.081 9797 1010 52.9 0.061 7070 3737 51.9 0.041 5959 6565 Downshift 47.4 0.028 99 8383 49.9 0.061 3737 3838 51.6 0.076 5656 1616 Table 9: Raw handoff by switch position in sharded, underspecified conversations, Claude pair. Early, middle, and late are structural switch positions. Later escalation retains more savings, while quality is mostly stable unless prior turns accumulate deliverable state. Later downshift generally improves quality while retaining less savings. Score is pass rate except for Data-to-text, which reports BLEU×100× 100. Task Policy Early Middle Late Score Cost QRec CSRet Score Cost QRec CSRet Score Cost QRec CSRet ($) (%) (%) ($) (%) (%) ($) (%) (%) Code LC-only 85.0 0.017 00 100100 85.0 0.017 00 100100 85.0 0.017 00 100100 HC-only 95.0 0.057 100100 00 95.0 0.057 100100 00 95.0 0.057 100100 00 Escalation 96.0 0.052 110110 1313 94.0 0.037 9090 4949 94.0 0.024 9090 8282 Downshift 90.0 0.024 5050 8282 94.0 0.037 9090 5050 96.0 0.052 110110 1313 Database LC-only 65.4 0.008 – 100100 65.4 0.008 – 100100 65.4 0.008 – 100100 HC-only 61.7 0.027 – 00 61.7 0.027 – 00 61.7 0.027 – 00 Escalation 64.5 0.025 – 1414 65.4 0.019 – 4444 66.4 0.014 – 7070 Downshift 60.7 0.012 – 8080 59.8 0.017 – 5252 64.5 0.024 – 1818 Actions LC-only 73.3 0.005 00 100100 73.3 0.005 00 100100 73.3 0.005 00 100100 HC-only 79.0 0.028 100100 00 79.0 0.028 100100 00 79.0 0.028 100100 00 Escalation 82.9 0.023 167167 2323 79.0 0.017 100100 5050 82.9 0.009 167167 8383 Downshift 73.3 0.011 00 7676 68.6 0.017 −83-83 4949 70.5 0.023 −50-50 2323 Math LC-only 92.2 0.005 – 100100 92.2 0.005 – 100100 92.2 0.005 – 100100 HC-only 94.2 0.024 – 00 94.2 0.024 – 00 94.2 0.024 – 00 Escalation 94.2 0.022 – 1010 92.2 0.018 – 3232 96.1 0.010 – 7676 Downshift 89.3 0.007 – 9191 91.3 0.011 – 6868 91.3 0.019 – 2424 Data-to-text LC-only 33.4 0.008 00 100100 33.4 0.008 00 100100 33.4 0.008 00 100100 HC-only 49.9 0.036 100100 00 49.9 0.036 100100 00 49.9 0.036 100100 00 Escalation 49.3 0.029 9696 2424 49.5 0.024 9898 4040 45.6 0.018 7474 6363 Downshift 32.9 0.009 −3-3 9595 32.7 0.023 −4-4 4545 37.0 0.030 2222 2222 Table 10: Raw handoff by switch position in sharded, underspecified conversations, GPT pair. Early, middle, and late are structural switch positions. Later escalation retains more savings, while quality is mostly stable unless prior turns accumulate deliverable state. Later downshift generally improves quality while retaining less savings. Score is pass rate except for Data-to-text, which reports BLEU×100× 100. Prefix model Difficulty N Switch step by prefix-run percentile p5 p10 p15 p25 p35 p45 p50 Haiku 4.5 Easy 194 25 30 34 37 42 46 50 Medium 261 35 41 45 51 58 63 66 Hard 45 48 54 58 63 77 86 94 Opus 4.7 Easy 194 5 7 8 10 12 14 14 Medium 261 8 12 13 16 18 21 23 Hard 45 18 19 20 24 30 33 37 GPT-5.6 Luna Easy 194 7 8 8 9 9 10 10 Medium 261 8 8 9 10 10 11 11 Hard 45 9 10 11 11 11 12 12 GPT-5.6 Sol Easy 194 9 10 10 11 12 13 13 Medium 261 10 11 11 13 13 15 15 Hard 45 12 13 14 15 16 18 19 Table 11: Difficulty-calibrated switch steps by prefix model. Steps are agent turns, corresponding to model API calls. Each threshold is the specified percentile of the prefix model’s single-model termination-step distribution within the given difficulty bucket.

Instant research discovery