Google Research proposes sequential auctions instead of a central orchestrator for multi-agent LLMs

A Google Research paper with collaborators replaces fixed workflow graphs for multi-agent LLM systems with sequential auctions, where agents bid for the right to execute each next step. The work shows convergence to globally optimal policies in MDPs under imperfect competition and compares 1st-price, 2nd-price and budget-aware Vickrey protocols.

• Agents bid for the next task step, so the workflow emerges from the market
• Three protocols compared: 1st-price, 2nd-price and budget-aware Vickrey
• If the winner fails, the next-highest bidder takes over the step
• Observability shifts to bid histories and market clearing prices