An arXiv preprint reports that a decentralized market for LLM agents posted its strongest reported benchmark scores in the setup that allowed subcontracting on GAIA tasks. The subcontracting version scored 57.6, 56.2, 54.8 and 48.4 in alpha order across four cost-sensitivity settings, averaging 54.2, compared with 52.2 without subcontracting.
The score balanced benchmark correctness against task execution cost, with alpha tested at 1, 2, 5 and 10; larger alpha values put more weight on cost. The paper reports statistically significant average improvements over every baseline at p < 0.05, but the statistical test and confidence intervals are not reported.
On the subset of tasks where subcontracting was invoked, the reported accuracy was 27.3% higher and the reported cost 64.1% lower. That comparison applies only to invoked tasks, a distinction that matters because subcontracting was enabled only for GAIA in the main setup.
A market built around hidden costs
The system, called AgentLance, treats coordination as a repeated labor market. Agents bid using execution costs that are private to them and keep self-maintained strategy notes; an allocator uses those bids and public reputation, then applies a VCG-style payment rule.
The evaluation used four randomly sampled 100-task pools, one for each benchmark. Each pool was split into 25 warmup tasks and 75 evaluation tasks, with the warmup portion excluded from reported performance.
The labor pool comprised GPT-5 nano, Grok 4.3, Gemini 3 Flash and DeepSeek-V3.2, while GPT-5 mini served as allocator. Four tasks were released per round, and the LLM-call methods were run three times with average scores reported.
The comparison set included fixed single-model policies, a best-model oracle for each alpha setting, centralized planner and router methods, and MarketBench direct and self-knowledge variants.
Allocation followed strengths and prices
In a preliminary two-agent test, allocation followed benchmark-specific strengths. Gemini's accuracy on SuperGPQA was 0.64 versus DeepSeek's 0.55, and Gemini received 89.2% of the task share; on GAIA, DeepSeek's accuracy was 0.27 versus Gemini's 0.20, and DeepSeek received 77.6%.
In a separate two-model cost-sensitivity test, GPT-5 nano's share rose from 3.3% at alpha 1 to 70.7% at alpha 100. Grok was more accurate, at 0.76 versus 0.61, but its average cost was $0.0045 compared with $0.0014 for GPT-5 nano.
The agents struggled to estimate their own costs
The clearest weakness was cost prediction. Across the models, correlations between direct execution-cost estimates and realized costs ranged from 0.091 to 0.299. Mean absolute percentage error (MAPE), the size of the percentage miss, ranged from 114% to 2,308%.
Historical cost information did not reliably resolve the problem: the highest observed correlation after it was used was 0.354. The share of bidding notes mentioning cost-truthful bidding rose from 47.9% at the beginning to 66.7% at the end of the market rounds, but this measures note content rather than verified bidding behavior; about one-third of notes still did not mention the strategy.
Useful signals, but a narrow test
Removing reflective notes was associated with an average score of 48.6%, compared with 52.2% with notes. Replacing the VCG-style critical payment with the winner's own bid was reported as 2.5% lower on average.
A joint cost-estimation and bidding-strategy condition showed positive score changes at every tested alpha: 1.5 points at alpha 1, 2.5 at alpha 2, 5.4 at alpha 5 and 7.7 at alpha 10. The interventions were controlled, but they were not reported as randomized experiments.
The limits are substantial: the experiment used four benchmark pools, four underlying model agents and one allocator, averaged three independent LLM-call runs, and reported no confidence intervals or run-level dispersion. Subcontracting was enabled only for GAIA, so the results do not establish how the mechanism would perform with larger agent pools, other allocators or deployed agent economies.
The supplied record identifies the work as an arXiv preprint; its peer-reviewed publication status is not reported.
Paper data and sources
Original title: Markets, Not Planners: Decentralized Orchestration of LLM Agents with Private Information
Authors: Xiao Liu, Haoyang Li, Songwei Li et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-24
DOI: Not available
Original paper · Full text