Agents that bid for their own work
AgentLance replaces the central planner with a labour market. Agents bid from private cost information, and work shifts to the cheaper ones.
2 minAutonomous AI Agents
Xiao Liu and six co-authors posted AgentLance on 24 August. It coordinates a group of language model agents without a planner: tasks are put out to bid, each agent bids from its own private execution cost, and the assignment falls out of the auction. The authors model it as a repeated labour market and allow hierarchical delegation, so an agent that wins work can put part of it back out.
The result
Across mathematical reasoning, code generation and knowledge-intensive tasks, the market matched agents to their specialisations and moved work toward the cheaper ones, beating centralised allocation. The authors also went looking for where the market failed, in inaccurate cost estimation and poor bidding strategy, and showed that fixing those lifted performance further.
The finding that should worry anyone shipping a planner
In a centralised allocator, inserting a single preference nearly doubled the favoured agent's share of tasks. One line, and the distribution of work moves that far.
That is the practical argument for the market design, and it is not really about efficiency. A central allocator built on a language model is a component whose routing policy can be rewritten by anything that reaches its context. A bidding mechanism moves the decision into a rule that does not read prose. Whether the market is faster is an empirical question; whether it is harder to steer by prompt is close to structural.
Retold from arXiv. This is a summary in our own words; follow the link for the original reporting.