Instead of a scheduler deciding which agent does what, market-based allocation lets agents bid for work — tasks carry prices, agents carry costs and capabilities, and allocation emerges from the auction. The contract-net protocol from 1980s distributed computing, rebuilt with LLM workers.

Central dispatch breaks down exactly where agent fleets get interesting: heterogeneous capabilities, dynamic loads, and tasks whose difficulty is unclear until someone tries. The contract-net protocol (Reid Smith, 1980) solved this with a market loop: a manager announces a task, capable nodes bid with their fitness and availability, the manager awards the contract, and the winner executes and reports. Pricing makes trade-offs explicit that schedulers usually bury — a worker that just failed three similar tasks can honestly discount its bid, and an overloaded specialist can price itself out rather than becoming a queue everyone waits on.

With LLM agents, bids become richer than utilization numbers: an agent can self-assess capability fit, quote expected token cost, and even attach confidence from analogous past tasks. Markets also compose with observability: realized-vs-quoted cost per agent is a calibration signal that improves future bidding — the market literally learns who is trustworthy.

The Engineering Reality

• Bids are self-reported: agents over- and under-estimate; normalize with reputation scores derived from realized outcomes, or the market rewards confidence instead of competence.
• Common currency required: mixing "I'm 80% suitable" with "I cost 4,000 tokens" needs an explicit utility function — pick one and publish it to all bidders.
• Anti-thrash rules: re-auctioning a task after every failure creates churn; constrain re-bidding rounds like any retry budget.
• Escalation path: some tasks must never float to the cheapest bidder — capability floors and safety-critical tags bound the market.

The honest framing: markets are not more efficient than good central schedulers at small scale — they are more honest. They surface costs, capabilities and uncertainty as first-class data instead of hiding them inside a dispatch heuristic, and at fleet scale, honest data is the scarce resource.