AgentLance replaces the central planner with a VCG auction, after showing one inserted preference nearly doubles a favored agent's task share

arXiv·medium signalXBlueskyLinkedInCopy link

Posted 24 August (arXiv 2608.23867), this work attacks centralized LLM-agent orchestration on three grounds: it bottlenecks as agent pools grow, it needs private information like each agent's execution cost, and it is manipulable, with a single inserted preference nearly doubling a favored agent's task share under a centralized LLM allocator. AgentLance is a repeated labor market where agents bid using private costs and self-maintained strategy notes, an allocator picks winners from bids plus public reputation records, and a VCG-style payment rule rewards cost-aware bidding, with winners able to decompose and subcontract complex work through the same mechanism. Across math reasoning, code generation, knowledge-intensive QA, and agentic tasks it beats single-model, centralized-orchestration, and market baselines, and shifts work toward cheaper agents as cost sensitivity rises.