Paper deep dive
Parallelism or Concession? Concurrency-Aware Procurement Negotiation for Agentic Commerce
Xiaolin Xu, Donghao Zhu
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Agentic buyers can cheaply fork a procurement task into many parallel negotiations, but concurrency is not free: every thread consumes resources, and simultaneous agreements create cancellation and commitment risk. We study a one-unit post-order sourcing problem with a single hard-deadline negotiation window, in which a planner jointly chooses the number of seller-facing negotiators and a common procurement price cap. The model combines a product-specific acceptance curve with fulfillment loss, per-thread cost, and excess-commitment cost. We establish three structural results. First, holding the per-thread acceptance target fixed, the marginal value of another negotiator decays geometrically, yielding a conditional concurrency threshold. Second, under a convex quantile curve, parallelism substitutes for concession: more concurrent negotiators imply a weakly lower per-thread acceptance target and price cap. Third, when prices are more dispersed, Agentic buyers benefit by searching harder for bargains, but suffer when they instead try to guarantee procurement by offering higher prices. We operationalize these results in the Concurrency-Aware Negotiation Optimizer (CANO), a deterministic optimizer that jointly determines the optimal negotiation concurrency and procurement price cap for an agentic procurement system. Across different analytic market configurations and extensive Monte Carlo, finite-data, non-Gaussian, and correlated-seller stress tests, CANO consistently outperforms common heuristic policies while validating the predicted structural properties.
Tags
Links
- Source: https://arxiv.org/abs/2610.06017v1
- Canonical: https://arxiv.org/abs/2610.06017v1
Trouble viewing inline? Open PDF directly →
Full Text
66,193 characters extracted from source content.
Expand or collapse full text
Parallelism or Concession? Concurrency-Aware Procurement Negotiation for Agentic Commerce Conference: The Eighth International Conference on Distributed Artificial Intelligence; November 29–December 2, 2026; Hong Kong, ChinaThe Eighth International Conference on Distributed Artificial Intelligence (DAI ’26), November 29–December 2, 2026, Hong Kong, ChinaCCS: Computing methodologies Multi-agent systemsCCS: Theory of computation Algorithmic game theory and mechanism designCCS: Applied computing Online shopping Xiaolin Xu Affiliation: School of Business, Nanjing University, 16 Jin Yin Street, Gulou District, Nanjing, Jiangsu, 210093, China email: xuxl@nju.edu.cn and Donghao Zhu Note: The authors are listed alphabetically by family name. Donghao Zhu is the lead and primary contributor to this work. Affiliation: Institute of Business Sciences, University of Tsukuba, 3-29-1 Otsuka, Bunkyo-ku, Tokyo, 112-0012, Japan Affiliation: University of Tokyo Market Design Center, 7-3-1 Hongo, Bunkyo-ku, Tokyo, 113-0033, Japan email: zhu.donghao.gt@u.tsukuba.ac.jp © , Abstract. Agentic buyers can cheaply fork a procurement task into many parallel negotiations, but concurrency is not free: every thread consumes resources, and simultaneous agreements create cancellation and commitment risk. We study a one-unit post-order sourcing problem with a single hard-deadline negotiation window, in which a planner jointly chooses the number of seller-facing negotiators and a common procurement price cap. The model combines a product-specific acceptance curve with fulfillment loss, per-thread cost, and excess-commitment cost. We establish three structural results. First, holding the per-thread acceptance target fixed, the marginal value of another negotiator decays geometrically, yielding a conditional concurrency threshold. Second, under a convex quantile curve, parallelism substitutes for concession: more concurrent negotiators imply a weakly lower per-thread acceptance target and price cap. Third, when prices are more dispersed, Agentic buyers benefit by searching harder for bargains, but suffer when they instead try to guarantee procurement by offering higher prices. We operationalize these results in the Concurrency-Aware Negotiation Optimizer (Cano), a deterministic optimizer that jointly determines the optimal negotiation concurrency and procurement price cap for an agentic procurement system. Across different analytic market configurations and extensive Monte Carlo, finite-data, non-Gaussian, and correlated-seller stress tests, Cano consistently outperforms common heuristic policies while validating the predicted structural properties. Keywords: automated negotiation, agentic commerce, concurrent negotiation, procurement, multi-agent orchestration, reservation price 1. Introduction Large language model (LLM) agents are beginning to search, bargain, and transact on behalf of users. A distinctive capability of such systems is cheap replication: a single procurement task can be delegated to multiple seller-facing subagents in parallel. Replication increases the chance of finding an acceptable seller while reducing the concession required in any individual negotiation. However, simply contacting every seller is rarely optimal. Each additional negotiation thread consumes tokens, time, and seller attention, while near-simultaneous agreements may leave the buyer with multiple commitments for a single purchase. This trade-off has attracted attention in both classical automated negotiation and recent language-agent systems. Prior work on concurrent negotiation studies coordination, tentative agreements, and decommitment (Nguyen and Jennings, 2005; An et al., 2011; Bagga et al., 2020), while recent benchmarks evaluate language agents in multi-seller markets with parallel execution (Zhu et al., 2025; Liu et al., 2026c). Nevertheless, existing approaches typically treat the number of negotiation threads or the buyer’s reservation price as fixed inputs, or learn an end-to-end dialogue policy. They do not isolate the fundamental orchestration decision that arises before any dialogue begins: How many negotiation agents should be launched, and what procurement price cap should they share? Those controls are coupled. Holding a cap fixed while adding threads improves fulfillment but also increases the chance of duplicate acceptance. Holding concurrency fixed while raising the cap improves each thread’s success but directly reduces transaction surplus. Optimizing either control in isolation can therefore recommend a high deal rate and still destroy value. The coupling is especially consequential for language agents because replication is cheap enough to make large m operationally plausible, while their seller-facing actions remain difficult to coordinate after several conversations converge at nearly the same time. We answer this question for one-unit post-order procurement. After accepting a customer order, a platform observes N eligible sellers and an estimated offer-acceptance curve F. It chooses m≤Nm≤ N concurrent negotiators and a common terminal cap b. All contacted sellers must be launched before any terminal response arrives: The delivery service-level agreement leaves time for one negotiation window, but not for a second serial attempt. A seller accepts with probability F(b)F(b). Failure incurs loss L, each thread costs κ, and each acceptance beyond the one fulfilled agreement incurs residual commitment cost δ. The last term captures races that remain even with a commitment manager: cancellation work, seller harm, reputational loss, or a legally binding tentative action. Market estimator N,F^N,\ F Cano planner m⋆,q⋆,b⋆m ,\ q ,\ b Parallel negotiation agents 1,…,m⋆1,…,m seller-facing agents offers constrained by pi≤b⋆p_i≤ b for each i≤m⋆i≤ m Commitment manager execute at most one transaction cancel or record the rest outcome log, auditing, and acceptance-curve updates Figure 1. Cano as the economic control layer of an agentic procurement system. The market estimator supplies the eligible-seller count N and estimated acceptance curve F F. The planner selects concurrency m⋆m , acceptance target q⋆q , and procurement cap b⋆b . Seller-facing agents negotiate in parallel subject to pit≤b⋆p_it≤ b , and the commitment manager executes at most one binding transaction. Logged outcomes support auditing and future acceptance-curve updates.A horizontal architecture diagram with four rounded blocks: market estimator, CANO planner, parallel negotiation agents, and commitment manager. Solid arrows indicate the forward control flow. A dashed curved arrow returns outcome logs, audit information, and acceptance-curve updates from the commitment manager to the market estimator. Our contribution is a small, auditable control problem rather than another end-to-end bargaining policy. • We formulate concurrency and concession jointly through a single expected-profit objective. The objective separates fulfillment value, language-agent operating cost, and excess-commitment exposure. • We characterize a fixed-target concurrency threshold, prove that parallelism substitutes for concession under explicit regularity conditions, and identify a sign reversal in the value of price dispersion. • We implement Cano and evaluate it over a complete 1944-configuration factorial sweep. The main comparisons are analytic rather than sample-contingent; one million simulated orders validate the expectation, and additional experiments study estimation error, non-Gaussian markets, and correlated seller prices. The results reveal two practical insights. First, as the number of sellers increases, Agentic buyers can rely more on parallel search and less on raising prices to secure procurement. Second, greater price dispersion creates more opportunities to find bargains, but also makes guaranteed procurement more expensive. 2. Background and System Framework 2.1. Related Work Automated and concurrent negotiation. Automated negotiation has long combined game-theoretic, heuristic, and argumentation-based approaches (Jennings et al., 2001). Parallel protocols and commitment models allow one agent to bargain with several counterparties while controlling delay and decommitment (Aknine et al., 2004; Nguyen and Jennings, 2005). An et al. (An et al., 2011) coordinate tentative agreements and concessions for multiple resources, and Bagga et al. (Bagga et al., 2020) learn one-to-many policies with deep reinforcement learning. The recent taxonomy of Florijn et al. (Florijn et al., 2025) emphasizes that one-to-many settings differ principally in how negotiation outcomes interact. Our setting is the unit-demand case: successes are substitutes, and at most one can be retained. Unlike dialogue-level strategies, we obtain a product-aware ex ante rule for both the number of dialogues and their shared reservation cap. Language-mediated bargaining. End-to-end negotiation dialogue models jointly learn strategy and language (Lewis et al., 2017); modular models instead decouple a strategic act from its linguistic realization (He et al., 2018). This separation has become especially relevant for LLM agents. NegotiationArena, buyer-side bargaining evaluations, and self-play studies document strategic variation and irrational behavior (Bianchi et al., 2024; Xia et al., 2024; Fu et al., 2023). Consumer-market experiments report overpayment and unfavorable transactions (Zhu et al., 2025), while verifiable-reward training improves budget adherence and surplus extraction (Liu et al., 2026b; Liu et al., 2026a). AgenticPay provides 111 tasks from bilateral to many-to-many negotiation and directly contrasts sequential and parallel modes (Liu et al., 2026c). Cano is complementary: it is an outer economic controller, not a language policy or benchmark. In particular, its cap can be enforced after an LLM proposes a structured price action. 2.2. Positioning Our unit of analysis is an orchestration decision, rather than an utterance, dialogue trajectory, or learned bargaining equilibrium. This creates two differences from the closest systems work. First, the policy is chosen before seller-facing text is generated and is expressed in quantities that an execution layer can enforce: a thread count and a terminal cap. Second, the objective values fulfillment, price, computation, and duplicate commitments jointly; agreement rate alone is not the target. The theory does not assert that one language model persuades better than another. It specifies the economic authority within which any such model may act and supplies comparative statics that dialogue benchmarks can test. Figure 1 illustrates the overall workflow of Cano. The market estimator learns the eligible seller count N and acceptance curve F F from historical offer–response data. Based on these estimates, the planner jointly determines the optimal concurrency m⋆m , acceptance target q⋆q , and procurement price cap b⋆b . The selected seller-facing agents negotiate in parallel under the shared price cap, while the commitment manager executes at most one transaction and records the remaining commitments. The resulting outcome log is fed back to the market estimator to improve future acceptance-curve estimation. 2.3. System Architecture At planning time, the estimator supplies a product-state snapshot, an eligible seller set, and an estimated acceptance curve F. The planner uses the corresponding quantile function C(q)=F−1(q)C(q)=F^-1(q) to translate a target per-thread acceptance probability into the corresponding procurement price cap. It then either returns “cancel” or a tuple (m⋆,q⋆,b⋆)(m ,q ,b ) together with a selected set of m⋆m sellers. During execution, each language agent may propose negotiation actions, but every proposed action is validated by the execution layer before being dispatched. In particular, the execution layer enforces pi≤b⋆and∑i=1m⋆xi≤1,p_i≤ b _i=1^m x_i≤ 1, where xi∈0,1x_i∈\0,1\ is a binary commitment indicator that equals 1 if seller i is selected for the final binding transaction, and 0 otherwise. The parameter δ represents the residual cost incurred when simultaneous negotiations cannot be perfectly coordinated because of external latency, seller-side finalization, or legally binding tentative acceptance. It models unavoidable implementation risk rather than relaxing the execution constraints. The outcome log records the estimator version, planned tuple, offers, acceptances, rejected actions, final commitment, and cancellations. These records support off-policy recalibration of F, cap-violation audits, and human review without exposing the language model to unilateral authority over payment. Thus Figure 1 separates a statistical component that may be uncertain, an optimization component that is deterministic given its inputs, and an execution component that is safety critical. The architecture is presented to clarify the system context rather than to optimize every component. In particular, this paper focuses exclusively on the economic control layer, the planner that determines negotiation concurrency and procurement price caps, while estimator design, language policy, and execution engineering are outside the scope of this work. 3. Procurement Model Table 1. Core notation and operational interpretation. Symbol Meaning N eligible sellers (market thickness) m concurrently launched negotiators b=C(q)b=C(q) common terminal procurement cap V net value of a fulfilled order L loss from cancellation or failed fulfillment κ cost per launched negotiation thread δ cost per acceptance beyond the first 3.1. Market and Decisions One accepted customer order yields net value V>0V>0 if fulfilled, after non-procurement costs. There are N sellers that already satisfy quality and delivery filters. Seller i’s minimum acceptable terminal price is Ri∼FR_i F, independently and identically distributed conditional on the product state. The planner starts m∈1,…,Nm∈\1,…,N\ threads and assigns the same terminal procurement cap b to every thread. It may instead cancel immediately, denoted m=0m=0, receiving payoff −L-L. We remark that, if procurement is expected to be uneconomical, the platform may immediately cancel the order and refund the customer instead of initiating negotiation. The cap is a walk-away price, not an opening offer. The terminal-offer abstraction charges b whenever at least one seller accepts. It is conservative when a multi-round protocol sometimes closes below its cap (Baarslag et al., 2016). A seller accepts if and only if Ri≤b<VR_i≤ b<V. We write q=F(b)q=F(b) and C(q)=F−1(q)C(q)=F^-1(q) where q is per-thread acceptance probability and C is the reservation-price quantile curve. “Price dispersion” below always refers to dispersion in the latent distribution of seller reservation prices RiR_i, rather than dispersion in posted prices. For the structural results, F is continuous and strictly increasing on the price interval containing the active policies, so the displayed inverse is well defined. The implementation can instead consume a bounded empirical quantile and uses global numerical search rather than the derivative results. Remark (timing and information). The customer order is already accepted when sourcing begins. The platform first applies non-price eligibility filters and freezes a product-state estimate of F; It then chooses m,bm,b, launches the threads simultaneously, observes terminal acceptances at the end of the sourcing window, and retains at most one transaction. There is no opportunity to launch a replacement thread after observing failure without missing the fulfillment deadline. Outcomes update future estimates but do not change F within the current order. After conditioning on observable seller and product characteristics through eligibility filtering and market estimation, the remaining reservation prices are assumed to be i.i.d. Shared unobserved shocks violate it and are evaluated separately in Section 5.3. We restrict attention to a common cap. Exchangeability motivates this parsimonious choice, and it reduces the execution interface to one auditable scalar, but we do not claim it is optimal among arbitrary asymmetric mechanisms. If seller-specific features support distinct conditional curves FiF_i, the planner should choose seller-specific caps and a subset jointly; that heterogeneous assignment problem is outside the structural claims here. Likewise, F describes the terminal response to a cap under a fixed interaction protocol. It is not identified by posted prices, and it may change if the dialogue policy, seller population, or product state changes. 3.2. Expected Objective For m≥1m≥ 1, the number of acceptances is K∼Binomial(m,q)K (m,q). Hence Sm(q)=ℙ(K≥1)=1−(1−q)mS_m(q)=P(K≥ 1)=1-(1-q)^m and (1) Dm(q)=[(K−1)+]=mq−Sm(q).D_m(q)=E[(K-1)^+]=mq-S_m(q). When K=0K=0, payoff is −L−κm-L-κ m. When K≥1K≥ 1, the buyer keeps one agreement and receives V−b−κm−δ(K−1)V-b-κ m-δ(K-1). Taking expectations gives Π(m,b)= (m,b)= (V−b)Sm(F(b))⏟Expected fulfilled-order value−L[1−Sm(F(b))]⏟Expected fulfillment loss (V-b)S_m(F(b))_Expected fulfilled-order value- L[1-S_m(F(b))]_Expected fulfillment loss (2) −κm⏟Thread cost−δmF(b)−Sm(F(b))⏟Expected excess-commitment cost. - κ m_Thread cost- δ\mF(b)-S_m(F(b))\_Expected excess-commitment cost. The objective (2) consists of four economically meaningful components. The first term is the expected surplus generated when procurement succeeds, namely the net order value multiplied by the fulfillment probability. The second term is the expected loss from procurement failure. The third term captures the operating cost of launching negotiation threads. The final term represents the expected cost of excess commitments caused by multiple simultaneous acceptances. In acceptance space this becomes the central objective: (3) Π(m,q)=[V+L+δ−C(q)]Sm(q)−L−κm−δmq. (m,q)=[V+L+δ-C(q)]S_m(q)-L-κ m-δ mq. In addition to profit and fulfillment, we report two complementary measures of commitment risk. The expected number of acceptances beyond the first is Dm(q)D_m(q), defined in (1). The probability that at least two threads accept is Bm(q)=1−(1−q)m−mq(1−q)m−1.B_m(q)=1-(1-q)^m-mq(1-q)^m-1. DmD_m prices the linear exposure in the objective, whereas BmB_m is easier to interpret as the frequency of orders requiring decommitment handling. Both are zero for m=1m=1, but neither is controlled by fulfillment probability alone. The planner maximizes Π over m∈1,…,Nm∈\1,…,N\ and q∈[0,1]q∈[0,1], then compares the best procurement policy with immediate cancellation Π(0)=−L (0)=-L. Appendix A.1 gives the full derivation and boundary conventions. 4. Structural Results and Cano 4.1. A Threshold for Concurrency Fix q. The value of adding a (m+1)(m+1)-st thread is (4) Δm(q)=[V+L+δ−C(q)]q(1−q)m−κ−δq. _m(q)=[V+L+δ-C(q)]q(1-q)^m-κ-δ q. Proposition 4.1 (Concurrency threshold). For fixed q∈(0,1)q∈(0,1): (1) If V+L+δ−C(q)>0V+L+δ-C(q)>0, then Δm+1(q)<Δm(q) _m+1(q)< _m(q) for all m≥1m≥ 1. (2) If V+L+δ−C(q)≤0V+L+δ-C(q)≤ 0, then Δm(q)≤0 _m(q)≤ 0 for all m≥1m≥ 1. (3) With ties resolved in favor of fewer threads, the optimizer of the fixed-q problem is mq⋆=minm∈1,…,N:m=N or Δm(q)≤0.m_q = \m∈\1,…,N\:m=N or _m(q)≤ 0 \. Proposition 4.1 establishes a threshold structure for concurrency. For a fixed per-thread acceptance target, the planner continues adding negotiators while the next thread has positive incremental value and stops once this value becomes non-positive. The comparative statics follow immediately from (4). Increasing V or L raises the incremental value of every additional thread, whereas increasing κ or δ lowers it. Consequently, the threshold weakly increases with V and L, and weakly decreases with κ and δ. The intuition is event-based. Thread m+1m+1 creates new fulfillment value only when all first m threads fail, an event with probability (1−q)m(1-q)^m. Operating and excess-commitment costs do not enjoy this geometric discount. The complete proof is in Appendix A.2. The threshold is directly computable without enumerating all subsets. If κ+δq>0κ+δ q>0, the stopping point is the smallest m≥1m≥ 1 such that (5) (V+L+δ−C(q))q(1−q)m≤κ+δq.(V+L+δ-C(q))q(1-q)^m≤κ+δ q. This expression separates a scarcity term, (V+L+δ−C(q))q(1−q)m(V+L+δ-C(q))q(1-q)^m, from a nonvanishing launch-and-commitment cost. It also shows why a large seller pool does not imply contacting every seller: N enlarges the feasible set, but it does not slow the geometric decay of the next thread’s contribution. If κ=δ=0κ=δ=0, every finite marginal is positive and the fixed-q subproblem selects N. 4.2. Parallelism Substitutes for Concession For fixed m, define Am(q)=1−(1−q)m(1−q)m−1=qm∑j=0m−1(1−q)−j.A_m(q)= 1-(1-q)^mm(1-q)^m-1= qm _j=0^m-1(1-q)^-j. We state the substitution result under conditions that rule out spurious nonconvex roots. Assumption 1 (Regular quantile). Let I=q∈(0,1):C(q)≤V+L+δI=\q∈(0,1):C(q)≤ V+L+δ\. On I, C is twice continuously differentiable, C′(q)>0C (q)>0, and C′(q)≥0C (q)≥ 0. For every fixed m under comparison, every global maximizer over q∈[0,1]q∈[0,1] lies in the interior of I. Assumption 1 imposes mild regularity conditions on the reservation-price quantile curve. Monotonicity (C′(q)>0C (q)>0) reflects the natural requirement that achieving a higher acceptance probability requires a weakly higher procurement cap. Convexity (C′(q)≥0C (q)≥ 0) rules out pathological quantile shapes and ensures that progressively higher acceptance targets become increasingly expensive. Finally, restricting the optimizer to the interior of I excludes degenerate corner solutions, such as always rejecting or always accepting, and ensures that the optimum is fully characterized by the first-order condition, allowing the comparative-statics analysis to focus on economically meaningful operating regimes. Theorem 4.2 (Parallelism–concession substitution). Under Assumption 1, the fixed-m optimizer qm⋆q_m is unique and satisfies (6) C(q)+C′(q)Am(q)+δ[(1−q)−(m−1)−1]=V+L.C(q)+C (q)A_m(q)+δ\! [(1-q)^-(m-1)-1 ]=V+L. Moreover, qm+1⋆≤qm⋆ and C(qm+1⋆)≤C(qm⋆).q_m+1 ≤ q_m \ and \ C(q_m+1 )≤ C(q_m ). Theorem 4.2 shows that the optimal bargaining policy depends on the available degree of parallelism. As more negotiation agents (larger m) are launched in parallel, the planner optimally lowers the target per-thread acceptance probability, which in turn implies a lower procurement cap. Consequently, parallelism substitutes for concession: additional negotiation threads allow each individual agent to bargain more aggressively rather than relying on higher prices to secure procurement. Therefore, an agentic buyer should adapt its bargaining policy to the available degree of parallelism instead of replicating one fixed strategy across different market conditions. Conversely, when API or seller-contact cost raises κ, the joint solution typically uses fewer threads and compensates with a higher cap. The proof in Appendix A.3 shows that both Am(q)A_m(q) and the commitment term in (6) grow with m, forcing the unique root leftward. Equation (6) is a marginal balance condition: the planner increases the procurement cap until the marginal gain from higher procurement reliability is exactly offset by the marginal procurement and duplicate-commitment costs. As more negotiation agents are launched in parallel, overall procurement becomes less dependent on any single negotiation thread, reducing the need to buy reliability through higher prices. Consequently, parallelism substitutes for concession: more parallel agents imply a lower optimal acceptance target and procurement cap for each individual negotiator. The thread cost κ does not appear in the first-order condition because it is independent of the procurement cap for fixed m. Instead, it affects the outer optimization that determines the optimal number of negotiation threads. Managerially, the theorem suggests that, when parallelism is inexpensive, agentic buyers should invest computational resources in broader parallel search rather than higher procurement prices. 4.3. When Dispersion Helps Consider a mean-preserving symmetric location-scale family, in which the mean reservation price remains fixed while the distribution becomes more or less dispersed symmetrically around the mean: Cσ(q)=μ+σz(q) and z(1−q)=−z(q),C_σ(q)=μ+σ z(q)\ and \ z(1-q)=-z(q), where σ controls the degree of price dispersion and symmetry implies that reservation prices spread equally above and below the mean. So z(q)z(q) is negative below the median and positive above it. The uniform family used for the main experiment has z(q)=3(2q−1) andRi∼U[μ−3σ,μ+3σ].z(q)= 3(2q-1)\ and\ R_i U [μ- 3σ,μ+ 3σ ]. Theorem 4.3 (Dispersion sign reversal). Fix m. Suppose its maximizing target is unique and interior. Let Πm⋆(σ) _m (σ) denote the resulting maximized profit. At differentiability points, dΠm⋆(σ)dσ=−z(qm⋆)Sm(qm⋆). d _m (σ)dσ=-z(q_m )S_m(q_m ). Hence dispersion raises optimal value when qm⋆<1/2q_m <1/2, lowers it when qm⋆>1/2q_m >1/2, and is locally neutral at qm⋆=1/2q_m =1/2. Parallel search makes the lower tail usable. In a thick market, where many eligible sellers are available, the buyer can achieve high overall fulfillment even when each individual negotiation has a relatively low acceptance probability (qm⋆<1/2q_m <1/2). Consequently, spreading reservation prices creates more opportunities to find low-price sellers. By contrast, in a thin market, where only a few eligible sellers are available, or when fulfillment reliability is critical, the planner must keep the target acceptance probability above the median. Although Theorem 4.3 is stated for a fixed concurrency level m, the same conclusion extends piecewise to the joint optimization problem. Specifically, whenever the optimal concurrency m⋆(σ)m (σ) remains constant over an interval of σ, the value function satisfies Π⋆(σ)=Πm⋆(σ) (σ)= _m (σ) and the dispersion result applies directly. Non-differentiability may occur only at switching points where the optimizer changes from one concurrency level to another, i.e., m⋆(σ−)≠m⋆(σ+)m (σ^-)≠ m (σ^+). The complete envelope argument is provided in Appendix A.4. The result is deliberately local to a location–scale family. It does not say that every mean-preserving spread is beneficial below the median, nor that an increase in observed listing-price variance has this effect. The operative object is the estimated terminal acceptance quantile. The sign result is useful because it converts a distributional change into a testable policy boundary: After re-optimization, the active target q⋆=.5q =.5 separates a lower-tail-search regime from a reliability-through-concession regime. 4.4. Cano Optimizer Cano enumerates m=1,…,Nm=1,…,N, solves the one-dimensional fixed-m problem, evaluates (3), and compares the best result with −L-L. Under Assumption 1, the left side of (6) is strictly increasing, so bisection with boundary checks returns a globally optimal policy to tolerance ε using O(Nlog(1/ε))O(N (1/ )) evaluations of the quantile function and its derivative. Empirical acceptance curves and mixtures need not satisfy the assumption. Our implementation therefore uses a bounded global scan in q, followed by golden-section refinement around the best cell. Unlike derivative-based root-finding methods, this implementation does not rely on a favorable initialization and therefore remains robust for empirical acceptance curves. This version costs O(NG+Nlog(1/ε))O(NG+N (1/ )) where G is the number of grid points in the initial coarse search. The first term corresponds to the global scan, while the second is due to the subsequent golden-section refinement. Complete implementation details, including pseudocode, correctness, boundary handling, and numerical validation, are provided in Appendix B. 5. Computational Evaluation 5.1. Design We ask three questions: whether the numerical optimizer reproduces the structural results; how costly common restricted policies are; and how sensitive the planner is to estimating F or misspecifying seller dependence. Parameters and policies. We normalize μ=1μ=1. The complete analytic factorial crosses N N ∈1,2,5,10,20,50, ∈\1,2,5,10,20,50\, σ σ ∈.05,.15,.30,.45, ∈\.05,.15,.30,.45\, V−μ V-μ ∈.1,.3,.5, ∈\.1,.3,.5\, L L ∈0,.1,.3, ∈\0,.1,.3\, κ κ ∈.001,.01,.03, ∈\.001,.01,.03\, δ δ ∈0,.02,.1, ∈\0,.02,.1\, for 19441944 configurations. We compare joint Cano with: (i) Single-agent Optimal (m=1m=1, optimize q), representing the traditional one-to-one negotiation paradigm without parallelism; (i) All-parallel (m=Nm=N, optimize q), representing the strategy of launching every available negotiation agent; (i) Mean-cap (b=μb=μ, optimize m), representing a simple market-heuristic pricing rule with a fixed procurement cap at the market mean; and (iv) Fixed-q (q=.8q=.8 per thread, optimize m), representing a service-level heuristic that fixes a conventional high per-thread acceptance target while adjusting only concurrency. Every policy may immediately cancel. Because these baselines restrict one or both decision dimensions, they quantify the value of joint optimization. For each market configuration, we define the regret of a restricted policy as Rpolicy=maxm∈0,…,Nq∈[0,1]ΠCano(m,q)−Πpolicy,R_policy= _ subarraycm∈\0,…,N\\\ q∈[0,1] subarray _ Cano(m,q)- _policy, where policy∈Single-agent,All-parallel,Mean-cap,Fixed-qpolicy∈\ Single-agent, All-parallel, Mean-cap, Fixed-q\ and ΠCano(m,q) _ Cano(m,q) denotes the expected profit achieved by the proposed joint optimizer. Regret therefore measures the opportunity cost of restricting the feasible decision space, and the informative outcome is its magnitude and location rather than the fact that it is nonnegative. We additionally report the ratio of aggregate surplus above −L-L, the share of cells within one percent of the oracle surplus (with a 0.010.01 scale floor near zero), and cancellation frequency. No confidence interval is attached to the factorial sweep because it exhausts the declared design rather than sampling configurations. Main comparisons use analytic expected values and therefore have no sampling interval. Reference runs use 2,049 q-values before refinement; The factorial and regime-map runs use 257, estimation and dependence runs use 513, and distributional and sensitivity runs use 1,025. These resolutions were selected through preliminary convergence checks, balancing numerical accuracy against computational cost; denser grids produced negligible changes in the resulting objective values while requiring substantially more evaluations. Empirical-quantile runs also include every interpolation knot. Sixteen unit tests cover the expectation, marginal identity, first-order condition, monotonicity, dispersion derivative, distribution inversion, cancellation, correlated extension, and agreement with a 100,001-point exhaustive grid. A one-million-order Monte Carlo run uses seed 20,260,803. Further details are in Appendix C. 5.2. Structural and Policy Results Policy adaptation across market regimes. Figure 2 (a) and (b) sweep the candidate concurrency level m=1,…,Nm=1,…,N. At each value of m, the planner holds m fixed and solves the one-dimensional problem qm⋆∈argmaxq∈[0,1]Π(m,q)q_m ∈ _q∈[0,1] (m,q). Panel (a) plots the resulting conditional optimum qm⋆q_m , whereas panel (b) expresses the same solution in price units through bm⋆=C(qm⋆)b_m =C(q_m ). Thus, panel (b) is the price-space representation of panel (a), rather than the result of a separate optimization over b. The vertical dashed line marks the jointly optimal concurrency m⋆m , obtained by comparing Π(m,qm⋆) (m,q_m ) across all candidate thread counts and with immediate cancellation. Figure 2. Structural predictions for V=1.4,μ=1,L=.2,κ=.01,δ=.05V=1.4,μ=1,L=.2,κ=.01,δ=.05. (a–b) With N=20,σ=.30N=20,σ=.30, the fixed-m acceptance target and cap fall as concurrency grows; the dashed line marks the joint optimum. (c) The joint policy over market thickness and dispersion. The white contour marks q⋆=.5q =.5, the sign boundary in Theorem 4.3.Three panels. The first plots an acceptance target decreasing with concurrent threads. The second plots a decreasing price cap. The third is a heat map in seller count and price dispersion, with higher thread counts in thicker and more dispersed markets and a white median contour. Panel (c) instead varies the market environment. Its horizontal axis is the number of eligible sellers N, and its vertical axis is reservation-price dispersion σ. At each point (N,σ)(N,σ), Cano jointly re-optimizes m and q; the color reports the resulting m⋆m , with darker colors indicating fewer threads and brighter colors indicating more. The stepwise color bands reflect the discreteness of m⋆m . The white contour is the set of market configurations for which the jointly optimized target satisfies q⋆=.5q =.5. The higher-dispersion side of the contour corresponds to q⋆<.5q <.5, whereas the lower-dispersion side corresponds to q⋆>.5q >.5. The point (N,σ)=(20,.30)(N,σ)=(20,.30) in panel (c) is exactly the reference market shown in panels (a) and (b): it yields (m⋆,q⋆,b⋆)=(13,0.161,0.648)(m ,q ,b )=(13,0.161,0.648) and expected profit 0.4650.465. Although the optimal per-thread acceptance target falls from q1⋆=.539q_1 =.539 to q⋆=0.161q =0.161, aggregate fulfillment rises to Sm⋆(q⋆)=1−(1−q⋆)m⋆≈.898S_m (q )=1-(1-q )^m ≈.898. Figure 2 translates the structural results into an operating map. Panels (a) and (b) show how the optimized per-thread acceptance target and procurement cap adjust as concurrency increases, while panel (c) identifies the market conditions under which broader parallel search becomes valuable. At low dispersion, the optimizer uses only a few relatively reliable threads because there is little lower-tail value to exploit. As N and σ increase, optimal concurrency generally rises but remains well below N; The downward-sloping q⋆=.5q =.5 contour further indicates that thicker markets can enter the lower-tail-search regime at lower levels of dispersion. Managerially, eligible sellers should therefore be treated as an option pool rather than as a launch target: Broader search should be paired with tighter price authority, and the buyer should choose between reliability through concession and reliability through parallel lower-tail search according to market thickness and the estimated reservation-price distribution. Two grouped bar charts over four market regimes. The left shows expected profit for five policies and the right shows fulfillment probability. CANO has the highest profit, especially in dispersed markets; some more reliable policies have lower profit. Figure 3. Representative markets with N∈3,20N∈\3,20\ and σ∈.08,.35σ∈\.08,.35\, holding V=1.4,L=.2,κ=.01,δ=.05V=1.4,L=.2,κ=.01,δ=.05. Joint optimization matters most in dispersed markets. Higher reliability is not free: Mean-cap and Fixed-q can fulfill more often while earning less.Two grouped bar charts over four market regimes. The left shows expected profit for five policies and the right shows fulfillment probability. CANO has the highest profit, especially in dispersed markets; some more reliable policies have lower profit. Profit–reliability trade-off. Figure 3 compares expected profit in panel (a) and system-level fulfillment probability in panel (b) across four representative (N,σ)(N,σ) regimes. Policy differences widen substantially as price dispersion increases. In the thick dispersed market, Cano earns 0.5190.519 at m=14,q=.151m=14,q=.151, compared with 0.0900.090 for Single-agent and 0.2690.269 for Mean-cap, even though All-parallel and Mean-cap attain slightly higher fulfillment. Together, the two panels show that higher reliability may be too costly when achieved through excessive price concession or unnecessary concurrency. Cano instead selects the profit-maximizing reliability level by balancing procurement price, fulfillment loss, thread cost, and excess-commitment exposure. Table 2. Full-factorial policy summary over 1,944 analytic configurations. Surplus ratio aggregates profit above immediate cancellation. Regret is in normalized profit units; “Near” is the share within 1% of Cano’s recoverable surplus; “Cancel” is the share choosing m=0m=0. Policy Surplus ratio Mean regret 90th-pct. regret Near (%) Cancel (%) Mean m Cano 1.000 0.000 0.000 100.0 0.0 7.07 Single-agent 0.614 0.160 0.425 28.8 0.0 1.00 All-parallel 0.857 0.059 0.234 62.8 10.9 10.46 Mean-cap 0.727 0.113 0.319 8.5 0.0 2.96 Fixed-q=.8q=.8 0.440 0.232 0.575 3.6 25.9 1.23 Costs of restricted orchestration. Table 2 summarizes performance over the full factorial and reveals two distinct costs of restricted orchestration. Single-agent, All-parallel, Mean-cap, and Fixed-q recover 61.4%61.4\%, 85.7%85.7\%, 72.7%72.7\%, and 44.0%44.0\%, respectively, of Cano’s recoverable surplus, defined as aggregate profit above the immediate cancellation payoff −L-L. All-parallel is within one percent of the joint solution in 62.8%62.8\% of configurations, but cancels in 10.9%10.9\% because launching all N threads can be worse than stopping, whereas Single-agent is within one percent in only 28.8%28.8\%. The underlying regret patterns clarify why. Single-agent leaves lower-tail search unused: its mean regret rises from 0.0160.016 at σ=.05σ=.05 to 0.3000.300 at σ=.45σ=.45, and from zero at N=1N=1 to 0.2520.252 at N=50N=50. All-parallel accesses the lower tail but pays for threads it does not need, with mean regret increasing from zero at N=1N=1 to 0.2110.211 at N=50N=50. The joint policy instead expands concurrency sublinearly: Averaged over all other factors, mean m⋆m rises from 1.001.00 at N=1N=1 to 17.0617.06 at N=50N=50, while the corresponding mean target falls from q⋆=.610q =.610 to .291.291. Aggregate fulfillment nevertheless increases from .610.610 to .888.888, showing that lower per-thread concession and higher system-level reliability can coexist. These comparisons quantify the opportunity cost of restricting the orchestration decision space; They do not claim that deterministic optimization outperforms an unconstrained learned policy. 5.3. Estimation and Dependence Stress Tests We next study how estimation error affects planning by replacing the true quantile function with an empirical quantile estimated from n∈25,50,100,250,1000n∈\25,50,100,250,1000\ historical reservation samples. For each sample size, we generate 100 independent datasets and evaluate the resulting policy under the true distribution. Mean oracle regret decreases from 0.06860.0686 (95% bootstrap interval [0.0528,0.0860][0.0528,0.0860]) at n=25n=25 to 0.00330.0033 ([0.0027,0.0041][0.0027,0.0041]) at n=1000n=1000, while mean acceptance-calibration error falls from .059.059 to .010.010. Thus, even an estimated acceptance curve is sufficient for good planning, although limited data can still lead to noticeably suboptimal concurrency decisions. This experiment intentionally represents a favorable estimation setting because it directly observes sellers’ latent reservation thresholds. In practice, historical negotiation logs contain only offer–response outcomes, so reservation thresholds are interval-censored and may also be affected by the policies that generated the data. Estimating F F in deployment therefore requires additional modeling assumptions. Our results isolate only the effects of finite-sample estimation error and integer policy decisions, leaving these identification issues aside. Table 3. Finite-data planning for N=20,μ=1,σ=.35,V=1.4,L=.2,κ=.01,δ=.05N=20,μ=1,σ=.35,V=1.4,L=.2,κ=.01,δ=.05. Each row uses 100 independent histories; intervals are percentile bootstrap intervals for mean oracle regret. Calibration is mean absolute acceptance error at the selected cap. n Mean regret 95% interval Mean m Calibration 25 .0686 [.0528, .0860] 12.56 .0591 50 .0251 [.0197, .0309] 12.95 .0337 100 .0220 [.0171, .0273] 13.53 .0306 250 .0092 [.0070, .0115] 14.06 .0173 1000 .0033 [.0027, .0041] 13.99 .0095 Table 3 summarizes the finite-data results. Because each sample size uses independently regenerated histories rather than nested datasets, the bootstrap intervals for n=50n=50 and n=100n=100 overlap, so their small difference should not be interpreted as a monotone improvement. Nevertheless, increasing the amount of historical data substantially reduces both planning regret and acceptance-calibration error. The reported calibration metric is local to the selected procurement cap and therefore does not assess the accuracy of the entire estimated acceptance curve. Estimation error is relatively modest compared with violating the independence assumption. We therefore introduce an exchangeable Gaussian copula while preserving every seller’s marginal acceptance distribution F, where ρ denotes the latent correlation governing shared seller shocks. Complete model details are provided in Appendix B.3. When ρ=.7ρ=.7, representing a strongly correlated market, the policy planned under independence sees expected profit fall from 0.5190.519 to 0.0420.042, whereas a correlation-aware planner recovers profit to 0.1690.169 by reducing concurrency from m=14m=14 to 44 and increasing the acceptance target from q=.151q=.151 to .363.363. As ρ increases from .1.1 to .7.7, the planner progressively selects fewer but more accommodating negotiation threads, showing that positive dependence weakens the diversification benefit of parallel negotiation: although each seller’s marginal acceptance probability remains q, fulfillment is no longer 1−(1−q)m1-(1-q)^m. Finally, the one-million-order Monte Carlo simulation differs from the analytic expected profit by only 0.00040, within two Monte Carlo standard errors, confirming the numerical correctness of the analytical model. Appendix D reports the complete robustness analyses, including alternative reservation-price distributions, copula-based dependence models, parameter sensitivity analyses, and the full performance curves. 5.4. What the Evaluation Establishes The evaluation establishes that the proposed economic controller behaves consistently with its theoretical predictions across the declared design space. The numerical optimizer reproduces the structural comparative statics, joint optimization consistently outperforms common restricted policies across market regimes, and the analytical objective closely matches Monte Carlo simulation. Additional finite-data, distributional, and dependence experiments demonstrate robustness to estimation error while identifying seller dependence as the principal source of model misspecification. These results validate the economic control layer rather than a complete agentic procurement system. They do not establish LLM instruction following, stationary seller behavior, or buyer profit as the appropriate social objective. Our contribution is an auditable optimization policy and execution boundary that future end-to-end agent evaluations can build upon, independent of any specific language model or API version. 6. Limitations, Responsible Use, and Conclusion 6.1. Limitations The model deliberately isolates a one-unit, single-cap procurement problem with conditionally i.i.d. sellers and a risk-neutral expected-profit objective. It does not model bundles, inventory, repeated interaction, endogenous seller strategies, or dynamic bargaining. . Acceptance curves must be learned from offer–response data rather than listing prices, and seller heterogeneity, correlation, and legal commitment differences are treated only through reduced-form extensions. We evaluate the economic controller, not LLM persuasion quality. The single negotiation window is essential. If failed negotiations could be retried immediately at negligible cost, sequential launch would weakly dominate parallel launch. Our results therefore apply only when all negotiations must be initiated before terminal responses arrive. Multi-window settings require dynamic batch-size decisions. The common-cap abstraction treats all prices below b as operationally equivalent. Real procurement may jointly optimize price, delivery, warranty, or other contract attributes, and multiround bargaining may settle below the cap. Extending the decision from one scalar to multiple attributes would require multidimensional acceptance models. Finally, the factorial grid is designed to cover representative market regimes rather than reflect their empirical frequency. Its normalization enables relative comparisons but does not calibrate κ or δ to currency units. Practical deployment therefore requires parameter estimation, periodic recalibration, and robust fallbacks under distribution shift. 6.2. Responsible Use Parallel agents can amplify spam, bargaining power asymmetry, deception, accidental commitments, and seller-contact externalities. Deployments should therefore rate-limit seller contact, disclose automation where required, distinguish tentative from binding actions, enforce price caps outside the language model, maintain auditable logs, and provide human oversight for high-stakes cases. Acceptance estimation should minimize personal data, while seller-level rate limits and cooldowns should constrain N before optimization, and κ should reflect seller-attention costs as well as computation. Applicable contract, consumer, privacy, and anti-discrimination law always take precedence over economic objectives. 6.3. Conclusion. Agentic replication changes the economics of negotiation. Its benefit is not simply a higher probability of agreement: Parallelism permits lower per-conversation concession, but only until geometric fulfillment gains are overtaken by thread and commitment costs at a fixed target. Cano jointly re-optimizes the target and turns that tradeoff into an auditable policy, and the dispersion sign reversal identifies when a varied market is an opportunity rather than a liability. These results support a design principle for agentic commerce: delegate language, but centrally optimize and enforce economic authority. References Aknine et al. (2004) Samir Aknine, Suzanne Pinson, and Melvin F. Shakun. 2004. An Extended Multi-Agent Negotiation Protocol. Autonomous Agents and Multi-Agent Systems 8, 1 (2004), 5–45. doi:10.1023/B:AGNT.0000009409.19387.f8 An et al. (2011) Bo An, Victor Lesser, and Kwang Mong Sim. 2011. Strategic Agents for Multi-Resource Negotiation. Autonomous Agents and Multi-Agent Systems 23, 1 (2011), 114–153. doi:10.1007/s10458-010-9137-2 Baarslag et al. (2016) Tim Baarslag, Mark JC Hendrikx, Koen V Hindriks, and Catholijn M Jonker. 2016. Learning About the Opponent in Automated Bilateral Negotiation: A Comprehensive Survey of Opponent Modeling Techniques. Autonomous Agents and Multi-Agent Systems 30, 5 (2016), 849–898. Bagga et al. (2020) Pallavi Bagga, Nicola Paoletti, Bedour Alrayes, and Kostas Stathis. 2020. A Deep Reinforcement Learning Approach to Concurrent Bilateral Negotiation. arXiv preprint arXiv:2001.11785 (2020). Bianchi et al. (2024) Federico Bianchi, Patrick John Chia, Mert Yuksekgonul, Jacopo Tagliabue, Dan Jurafsky, and James Zou. 2024. How Well Can LLMs Negotiate? Negotiationarena Platform and Analysis. arXiv preprint arXiv:2402.05863 (2024). Florijn et al. (2025) Tamara C. P. Florijn, Pinar Yolum, and Tim Baarslag. 2025. A Survey on One-to-Many Negotiation: A Taxonomy of Interdependency. In Proceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence. 10436–10444. doi:10.24963/ijcai.2025/1159 Fu et al. (2023) Yao Fu, Hao Peng, Tushar Khot, and Mirella Lapata. 2023. Improving Language Model Negotiation with Self-play and In-context Learning from AI Feedback. arXiv preprint arXiv:2305.10142 (2023). He et al. (2018) He He, Derek Chen, Anusha Balakrishnan, and Percy Liang. 2018. Decoupling Strategy and Generation in Negotiation Dialogues. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics, 2333–2343. doi:10.18653/v1/D18-1256 Jennings et al. (2001) Nicholas R Jennings, Peyman Faratin, Alessio R Lomuscio, Simon Parsons, Carles Sierra, and Michael Wooldridge. 2001. Automated Negotiation: Prospects, Methods and Challenges. International Journal of Group Decision and Negotiation 10, 2 (2001), 199–215. Lewis et al. (2017) Mike Lewis, Denis Yarats, Yann N. Dauphin, Devi Parikh, and Dhruv Batra. 2017. Deal or No Deal? End-to-End Learning of Negotiation Dialogues. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics, 2443–2453. doi:10.18653/v1/D17-1259 Liu et al. (2026a) Shuze Daniel Liu, Claire Chen, Jiabao Sean Xiao, Xin Chen, and David Simchi-Levi. 2026a. Strategic Bargaining in Multi-Buyer Markets: Reinforcement Learning from Verifiable Rewards for LLM Negotiations. arXiv preprint arXiv:2607.05863 (2026). Liu et al. (2026b) Shuze Daniel Liu, Claire Chen, Jiabao Sean Xiao, Lei Lei, Yuheng Zhang, Yisong Yue, and David Simchi-Levi. 2026b. Instructing LLMs to Negotiate Using Reinforcement Learning with Verifiable Rewards. arXiv preprint arXiv:2604.09855 (2026). Liu et al. (2026c) Xianyang Liu, Shangding Gu, and Dawn Song. 2026c. Agenticpay: A Multi-agent LLM Negotiation System for Buyer-Seller Transactions. arXiv preprint arXiv:2602.06008 (2026). Nguyen and Jennings (2005) Thuc Duong Nguyen and Nicholas R Jennings. 2005. Managing Commitments in Multiple Concurrent Negotiations. Electronic Commerce Research and Applications 4, 4 (2005), 362–376. Xia et al. (2024) Tian Xia, Zhiwei He, Tong Ren, Yibo Miao, Zhuosheng Zhang, Yang Yang, and Rui Wang. 2024. Measuring Bargaining Abilities of LLMs: A Benchmark and A Buyer-Enhancement Method. In Findings of the Association for Computational Linguistics: ACL 2024. 3579–3602. Zhu et al. (2025) Shenzhe Zhu, Jiao Sun, Yi Nian, Tobin South, Alex Pentland, and Jiaxin Pei. 2025. The Automated But Risky Game: Modeling and Benchmarking Agent-to-Agent Negotiations and Transactions in Consumer Markets. arXiv preprint arXiv:2506.00073 (2025). Appendix A Derivations and Complete Proofs A.1. Expected Objective For m≥1m≥ 1, let K=∑i=1mRi≤bK= _i=1^m1\R_i≤ b\. Conditional on the product state, K∼Binomial(m,q)K (m,q), where q=F(b)q=F(b). The realized payoff is U(K)=−L−κm,K=0,V−b−κm−δ(K−1),K≥1.U(K)= cases-L-κ m,&K=0,\\ V-b-κ m-δ(K-1),&K≥ 1. cases The excess-commitment identity follows pointwise from (K−1)+=K−K≥1(K-1)^+=K-1\K≥ 1\: [(K−1)+]=[K]−ℙ(K≥1)=mq−Sm(q).E[(K-1)^+]=E[K]-P(K≥ 1)=mq-S_m(q). Therefore [U(K)] [U(K)] =(V−b)Sm(q)−L[1−Sm(q)]−κm =(V-b)S_m(q)-L[1-S_m(q)]-κ m −δ[mq−Sm(q)] -δ[mq-S_m(q)] =[V+L+δ−b]Sm(q)−L−κm−δmq. =[V+L+δ-b]S_m(q)-L-κ m-δ mq. Substituting b=C(q)b=C(q) yields (3). The convention m=0m=0 represents immediate cancellation after the customer order has been accepted, so Π(0)=−L (0)=-L. For m≥1,q=0m≥ 1,q=0, procurement fails surely and Π(m,0)=−L−κm (m,0)=-L-κ m, weakly worse than cancellation. For q=1q=1 when C(1)<∞C(1)<∞, every launched thread accepts and Π(m,1)=V−C(1)−κm−δ(m−1) (m,1)=V-C(1)-κ m-δ(m-1); hence m=1m=1 dominates larger concurrency at this boundary. For unbounded support, no finite cap attains q=1q=1; optimization is over q∈[0,1)q∈[0,1), with the upper boundary interpreted by a limit. ■ A.2. Proof of Proposition 4.1 Since Sm+1(q)−Sm(q)=q(1−q)mS_m+1(q)-S_m(q)=q(1-q)^m and the expected acceptance count rises by q, subtracting Π(m,q) (m,q) from Π(m+1,q) (m+1,q) gives (4). Write G(q)=V+L+δ−C(q)G(q)=V+L+δ-C(q). If G(q)>0G(q)>0, then Δm+1(q)−Δm(q)=−G(q)q2(1−q)m<0. _m+1(q)- _m(q)=-G(q)q^2(1-q)^m<0. If G(q)≤0G(q)≤ 0, the first term of (4) is nonpositive, so Δm(q)≤−κ−δq≤0 _m(q)≤-κ-δ q≤ 0. Thus, over positive integers, profit rises while Δm(q)>0 _m(q)>0, is flat when Δm(q)=0 _m(q)=0, and falls after the marginal becomes negative. Resolving a zero marginal toward the smaller thread count gives the smallest m with Δm(q)≤0 _m(q)≤ 0; when no such m<Nm<N exists, the solution is N. For comparative statics, when G(q)>0G(q)>0, ∂Δm∂V=∂Δm∂L=q(1−q)m>0,∂Δm∂κ=−1, ∂ _m∂ V= ∂ _m∂ L=q(1-q)^m>0, ∂ _m∂κ=-1, and ∂Δm∂δ=q(1−q)m−q=−q[1−(1−q)m]≤0. ∂ _m∂δ=q(1-q)^m-q=-q[1-(1-q)^m]≤ 0. Moving every marginal in one direction moves its crossing threshold weakly in the same direction. N only truncates the resulting threshold. ■ A.3. Proof of Theorem 4.2 For compactness, set B(q)=V+L+δ−C(q).B(q)=V+L+δ-C(q). The first two derivatives of the fixed-m objective are Πm′(q) _m (q) =−C′(q)Sm(q)+B(q)Sm′(q)−δm, =-C (q)S_m(q)+B(q)S_m (q)-δ m, Πm′(q) _m (q) =−C′(q)Sm(q)−2C′(q)Sm′(q)+B(q)Sm′(q). =-C (q)S_m(q)-2C (q)S_m (q)+B(q)S_m (q). On the interval in Assumption 1, Sm′(q)=m(1−q)m−1>0S_m (q)=m(1-q)^m-1>0, Sm′(q)≤0S_m (q)≤ 0, B(q)≥0B(q)≥ 0, C′(q)>0C (q)>0, and C′(q)≥0C (q)≥ 0. Hence Πm′(q)<0 _m (q)<0, proving strict concavity and uniqueness. At an interior optimum, Πm′(q)=0 _m (q)=0. Dividing by Sm′(q)=m(1−q)m−1S_m (q)=m(1-q)^m-1 gives V+L+δ−C(q)=C′(q)Am(q)+δ(1−q)−(m−1).V+L+δ-C(q)=C (q)A_m(q)+δ(1-q)^-(m-1). Rearrangement yields (6). It remains to compare roots. Let Hm(q)=C(q)+C′(q)Am(q)+δ[(1−q)−(m−1)−1].H_m(q)=C(q)+C (q)A_m(q)+δ[(1-q)^-(m-1)-1]. For fixed q∈(0,1)q∈(0,1), write r=1−qr=1-q. Then Am(q)=qm∑j=0m−1r−j.A_m(q)= qm _j=0^m-1r^-j. The sequence 1,r−1,r−2,…1,r^-1,r^-2,… is strictly increasing, so appending its next term raises its running average: Am+1(q)>Am(q)A_m+1(q)>A_m(q). The commitment term is also nondecreasing in m. Thus Hm+1(q)≥Hm(q)H_m+1(q)≥ H_m(q). For fixed m, Am(q)A_m(q) is strictly increasing because both q and every r−jr^-j are nondecreasing and the j=0j=0 term makes the product strict. Together with C′>0,C′≥0C >0,C ≥ 0, this makes Hm(q)H_m(q) strictly increasing. Since both optima solve Hm(q)=V+LH_m(q)=V+L, qm+1⋆≤qm⋆.q_m+1 ≤ q_m . Finally C is increasing, so the same inequality holds for the caps. ■ A.4. Proof of Theorem 4.3 For fixed m, only the quantile curve in (3) depends on σ. At a unique optimum, the envelope theorem gives dΠm⋆(σ)dσ=−∂Cσ(qm⋆)∂σSm(qm⋆)=−z(qm⋆)Sm(qm⋆). d _m (σ)dσ=- ∂ C_σ(q_m )∂σS_m(q_m )=-z(q_m )S_m(q_m ). Because Sm(q)>0S_m(q)>0 for q>0q>0 and symmetry implies z(q)<0z(q)<0 below 1/21/2, z(1/2)=0z(1/2)=0, and z(q)>0z(q)>0 above 1/21/2, the three sign cases follow. If a positive joint optimizer over m is unique in a neighborhood of σ and its fixed-m solution remains unique and interior, the same envelope derivative applies to the active m. At an integer switch, the upper envelope of the fixed-m value functions is continuous but need not be differentiable. ■ Appendix B Cano Algorithm and Extensions B.1. Planner Output and Failure Behavior The planner receives a market instance containing (N,V,L,κ,δ)(N,V,L,κ,δ) together with the estimated quantile function C(q)C(q). It returns either a cancellation decision or an optimal policy tuple (m⋆,q⋆,b⋆),(m ,q ,b ), while fulfillment and commitment metrics are evaluated separately by a deterministic evaluator. The implementation first performs a bounded global search over q∈[10−6,1−10−6]q∈[10^-6,1-10^-6], followed by golden-section refinement around the best candidate. For finite-support distributions, the boundary point q=1q=1 is additionally checked. The resulting policy is finally compared with immediate cancellation, and cancellation is selected whenever every positive-concurrency policy yields a lower expected value than −L-L. On a reference uniform instance, the implementation matches an exhaustive 100,001-point grid search to within 10−810^-8. This serves as a regression check rather than a universal optimality guarantee. Theoretical global optimality remains guaranteed only under Assumption 1. Implementation details are provided in the following. B.2. Algorithm For each m∈1,…,Nm∈\1,…,N\, Cano performs: (1) Construct qg=q¯+g(q¯−q¯)/(G−1)q_g= q+g( q- q)/(G-1) for g=0,…,G−1g=0,…,G-1. (2) Evaluate Π(m,qg) (m,q_g) and retain the best cell and both neighbors. (3) Refine that bracket by golden-section maximization to tolerance ε , also checking global endpoints. (4) Convert the best qmq_m to bm=C(qm)b_m=C(q_m) and record its value. It returns the best recorded positive-m policy unless cancellation −L-L is larger. Under Assumption 1, Step 1 may be replaced by boundary derivative checks and bisection of (6). Strict concavity proves global correctness, while bisection needs O(log(1/ε))O( (1/ )) evaluations for each m. The global-scan implementation is intentionally more conservative for empirical curves. B.3. Correlated Sellers The i.i.d. theory does not apply when seller reservation prices share shocks. For the stress test, let latent acceptance variables follow an exchangeable Gaussian copula: Zi=ρZ0+1−ρϵi,Z_i= ρZ_0+ 1-ρ _i, where all normal factors are independent and a seller accepts when Zi≤Φ−1(q)Z_i≤ ^-1(q). Conditional on Z0=zZ_0=z, acceptance probability is pq(z)=Φ(Φ−1(q)−ρz1−ρ).p_q(z)= \! ( ^-1(q)- ρz 1-ρ ). Consequently, Smρ(q)=1−Z0[(1−pq(Z0))m].S_m^ρ(q)=1-E_Z_0[(1-p_q(Z_0))^m]. Marginal acceptance remains q, so [K]=mqE[K]=mq and the objective retains the form Πρ(m,q)=[V+L+δ−C(q)]Smρ(q)−L−κm−δmq. _ρ(m,q)=[V+L+δ-C(q)]S_m^ρ(q)-L-κ m-δ mq. We evaluate the one-dimensional expectation with 32-node Gauss–Hermite quadrature, enumerate m, and maximize over a 513-point q-grid. Positive dependence lowers the diversification benefit of concurrency, explaining the shift toward fewer, more accommodating threads. Three panels. Mean oracle regret falls from about 0.069 with 25 observed threshold samples to 0.003 with 1,000. As the tested latent Gaussian correlation rises from 0 to 0.7, expected profit falls; at every positive tested correlation, the correlation-aware line is above the independence-planned line. Hatched grouped bars compare five policies under four mean-one, standard-deviation-about-0.30 distribution shapes. Figure 4. Stress tests. (a) Mean regret and 95% bootstrap interval over 100 finite-data replicates. (b) Correlation-aware planning partially recovers profit lost under shared seller shocks. (c) Policy profit under uniform, truncated-normal, lognormal, and two-component reservation-price distributions.Three panels. Mean oracle regret falls from about 0.069 with 25 observed threshold samples to 0.003 with 1,000. As the tested latent Gaussian correlation rises from 0 to 0.7, expected profit falls; at every positive tested correlation, the correlation-aware line is above the independence-planned line. Hatched grouped bars compare five policies under four mean-one, standard-deviation-about-0.30 distribution shapes. Appendix C Additional Experimental Detail C.1. Full Protocol All results are generated by one Python script using NumPy and Matplotlib. The main seed is 20,260,803. Analytic rows, figures, and LaTeX result macros are written from the same run. The implementation contains separate modules for distributions, the profit model, optimization, correlation, and simulation. No proprietary data, model API, or human-subject data are used. Table 4. Factorial design for the 19441944 analytic markets. Parameter Values N 1,2,5,10,20,501,2,5,10,20,50 σ/μσ/μ .05,.15,.30,.45.05,.15,.30,.45 V−μV-μ .10,.30,.50.10,.30,.50 L 0,.10,.300,.10,.30 κ .001,.01,.03.001,.01,.03 δ 0,.02,.100,.02,.10 The primary metrics are expected profit, fulfillment probability, cap, concurrency, expected excess commitments mq−Sm(q)mq-S_m(q), duplicate probability ℙ(K≥2)=1−(1−q)m−mq(1−q)m−1,P(K≥ 2)=1-(1-q)^m-mq(1-q)^m-1, and absolute regret. Fixed-q=.7q=.7 and .9.9 sensitivities, as well as one-factor sweeps in V,L,κ,δV,L,κ,δ, are included in the raw output. C.2. Monte Carlo Validation For N=20,μ=1,σ=.35,V=1.4,L=.2,κ=.01,δ=.05N=20,μ=1,σ=.35,V=1.4,L=.2,κ=.01,δ=.05, Cano selects m=14,q=.15097,b=.57682m=14,q=.15097,b=.57682. Analytic expected profit is 0.5189560.518956. Across one million independent orders, simulated mean profit is 0.5193540.519354 with standard error 0.0002930.000293; absolute error is 0.0003970.000397. Simulated fulfillment and duplicate probabilities are 0.899240.89924 and 0.646610.64661. Appendix D Robustness Results D.1. Distribution Shape Table 5 reports the policies underlying Figure 4(c). All four distributions have mean one and standard deviation approximately 0.300.30. The truncated normal is symmetric at three underlying standard deviations. The two-component market contains a low-price segment and a smaller high-price segment; its quantile is represented by 100,000 fixed-seed draws and captures unobserved seller-type heterogeneity. Because some shapes violate the convex-quantile condition globally, the bounded grid scan followed by local refinement is used throughout. Table 5. Distribution robustness: m, q, and expected profit. Uniform Trunc. normal Policy m,qm,q Profit m,qm,q Profit Cano 13,.16113,.161 .465 15,.13415,.134 .421 Single 1,.5391,.539 .092 1,.6271,.627 .105 All 20,.11720,.117 .444 20,.10620,.106 .412 Mean-cap 4,.5004,.500 .269 4,.5004,.500 .269 Fixed-.8 2,.8002,.800 .025 2,.8002,.800 .079 Lognormal Mixture Policy m,qm,q Profit m,qm,q Profit Cano 11,.18711,.187 .406 5,.4295,.429 .477 Single 1,.6411,.641 .133 1,.5401,.540 .215 All 20,.11520,.115 .377 20,.13320,.133 .360 Mean-cap 3,.5583,.558 .280 3,.5503,.550 .278 Fixed-.8 2,.8002,.800 .107 2,.8002,.800 .003 D.2. Parameter Comparative Statics The numerical one-factor sweeps agree with Proposition 4.1. In the reference uniform market, raising κ from .001 to .06 reduces m⋆m from 20 to 4 and raises q⋆q from .117 to .325. Raising δ from 0 to .20 reduces m⋆m from 15 to 9 and expected excess commitments from 1.659 to .637. Raising L from 0 to .60 increases concurrency from 12 to 14 and fulfillment from .869 to .929. Because L also appears in the failure payoff, reported profit need not increase with L; the policy nevertheless becomes more reliable as predicted. D.3. Scope of the Agentic Evaluation The experiment targets the economic control layer shown in Figure 1. It samples seller acceptance and commitment outcomes but does not call an LLM, measure rhetoric, or compare prompts. Accordingly, no claim in the paper attributes profit improvements to language generation. A future controlled LLM evaluation should pair identical market seeds, seller messages, and model snapshots; distinguish attempted from executed cap violations; and report token use, latency, and prevented duplicate commitments. The current results establish the policy that such a language layer should be constrained to execute.