As LLM agents proliferate, built by different parties and with different capabilities and costs, orchestrating them is more like assembling labor across the economy than a computer calling a subroutine. Existing orchestration is typically centralized, with a single planner assigning every task, but this creates a bottleneck as agent pools grow, requires private information (e.g., agents' execution costs), and can easily be manipulated, such that a single inserted preference nearly doubles a favored agent's task share under a centralized LLM allocator. We introduce AgentLance, a repeated labor market in which agents bid on tasks using their private costs and self-maintained strategy notes, an allocator selects winners from bids and public reputation records, and a VCG-style payment rule rewards cost-aware bidding. Complex tasks are handled by hierarchical delegation: winning agents can decompose work and subcontract it through the same mechanism. Across mathematical reasoning, code generation, knowledge-intensive QA, and agentic tasks, AgentLance matches agents to their specializations, shifts work toward cheaper agents as cost sensitivity rises, and consistently outperforms single-model, centralized-orchestration, and market baselines. Diagnosing market failures, including inaccurate cost self-estimation and sub-optimal bidding, then correcting them in controlled experiments yields further gains, charting a path toward more efficient agent economies.
Figures & tables
Figure 1: Overview and motivation of AgentLance . (a) Agents bid on tasks, and selected agents may collaborate through subcontracting. Market outcomes update their strategies and reputations. (b) Inserted preferences can substantially bias the allocations of a centralized LLM planner.
Figure 2: Illustrative market behaviors. (a) The market assigns tasks to agents matching their specializations. (b) As cost sensitivity increases, more tasks are allocated to the cheaper model.
Method
Score (%) = Acc - α× Cost
α=1
α=2
α=5
α=10
Average
Single Model
GPT-5 nano
44.6
44.1
42.8
40.6
43.0
Grok 4.3
56.3
54.5
49.3
40.5
50.1
Gemini 3 Flash
56.5
52.9
42.3
24.6
44.1
DeepSeek-V3.2
52.1
50.1
44.3
34.7
45.3
Table 1: Performance comparison under different values of α . The best is in bold, and the second-best is underlined.
Method
Score (%) = Acc - α× Cost
α=1
α=2
α=5
α=10
Avg.
AgentLance w/o subcontract
56.3
55.3
51.6
45.6
52.2
w/o Reflective Notes
55.3
51.7
50.9
36.5
48.6
w/o VCG-style Payment
50.7
55.1
45.6
47.5
49.7
Table 2: Ablation study under different values of α . The best performance in each column is in bold.
Figure 3: Limitations and adaptation of market participants. Left: Agents exhibit low correlations and high errors when estimating execution costs. Right: Cost-truthful strategies generally become more prevalent over market rounds.
Cost
Strategy
Δ ( α=1 )
Δ ( α=2 )
Δ ( α=5 )
Δ ( α=10 )
✓
×
-1.0
-0.9
+1.9
+4.8
×
✓
+1.6
-2.2
-1.1
-11.4
✓
✓
+1.5
+2.5
+5.4
+7.7
Table 3: Score changes under control settings. Δ is computed as control minus original AgentLance (%).
Figure 4: Themes in agents’ reflective notes.
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Theme
Example
Task Fit
For factual lookup tasks (for example book chapter salients or author identifications), mandate explicit deliverables with documented data sources and provenance. If required data or sources are not accessible, request access prior to bidding or pass. Ensure outputs include provenance and templates/logs to support consistency across bids.
Success/Reputation
Reputation as Currency: p_hat is your most valuable asset. A failed task is more damaging than an empty round. Volume is secondary to accuracy.
Auction Positioning
Technical Bid Floor: For specialized STEM, bids below $0.015 are likely to lose.
Cost/Payment
Accurately estimate task costs before bidding; winning unprofitable tasks (where cost exceeds payment) leads to net losses regardless of success rate. However, avoid over-conservative multipliers (e.g., 100-200x) if historical data shows costs are consistently negligible for certain task types—calibrate estimates based on observed actual costs.
Risk Abstention
Escalation discipline: if a bid seems misaligned with effort or uncertainty is high, escalate or pass rather than bid; avoid forcing high-risk bids.
Appendix
Table C.1: Example snippets from agents’ notes.
Figure D.1 : Cumulative profit of individual agents in AgentLance w. subcontract across different values of α .
Enhancing the reasoning capabilities of large language model (LLM) agents requires effective orchestration of diverse expert models and tools. However, existing frameworks typically call APIs, based on coarse-grained matching between tasks and the functions of expert models or tools, while overlooking critical factors such as performance variability and cost efficiency among functionally similar alternatives. To address this, we propose Agora, a framework that uses a confidence-calibrated auction to dynamically allocate tasks to expert models and tools. By treating reasoning steps as tradeable items, Agora bases allocation on calibrated competence rather than raw confidence. Across five main benchmarks, Agora improves or remains competitive with single-model, routing, and cascade baselines under matched candidate pools.
Markets are a promising way to coordinate AI agent activity for similar reasons to those used to justify markets more broadly. In order to effectively participate in markets, agents need to have informative signals of their own ability to successfully complete a task and the cost of doing so. We propose MarketBench, a benchmark for assessing whether AI agents have these capabilities. We use a 93-task subset of SWE-bench Lite, a software engineering benchmark, with six recently released LLMs as a demonstration. These LLMs are miscalibrated on both success probability and token usage, and auctions built from these self-reports diverge from a full-information allocation. A follow-up intervention where we add information about capabilities from prior experiments to the context improves calibration, but only modestly narrows the gap to a full-information benchmark. We also document the performance of a market-based scaffolding with these LLMs. Our results point to self-assessment as a key bottleneck for market-style coordination of AI agents.
Andrey Fradkin, Rohit Krishnan
Boston University and the MIT Initiative on the Digital Economy. · Independent Researcher.
Agentic commerce is moving from concept to deployed infrastructure: payment networks, retailers, and AI platforms are setting the stage for agents to transact on behalf of merchants and consumers. Yet whether the LLMs behind these agents can price competently in real markets, where customer preferences are hidden, competitors adapt in real time, and demand can shift without warning, has not been systematically tested. We introduce Bazaar, a dynamic sealed-bid benchmark for multi-attribute auction under these conditions. Despite its dynamics, the benchmark is grounded in closed-form customer utilities, enabling exact evaluation. Across 11 frontier LLMs from four providers, the leading agents on customer acquisition (e.g. Gemini 3.1 Pro) are often not the leading agents on profit (e.g. Opus 4.6). The ranking shifts again under demand shocks: agents that learned fastest pre-shock are typically the slowest to revise their beliefs afterwards, while Gemini 3.1 Pro recovers fastest despite not leading on profit. However, even the strongest agent captures less than a third of hindsight-optimal profit, suggesting current LLMs are progressing in agentic commerce but leave substantial headroom.