Key Takeaway
Coordinating many small AI agents through a short-plan bidding system cuts reliance on the largest model by half and reduces overall cost by about a third while improving final answers.
ON THIS PAGE
Key Findings
Small models alone struggle as tasks get deeper and more complex, especially on long searches and coding challenges. Letting agents bid short strategic plans and scoring those plans with a cost-versus-value rule routes work more intelligently than simple description-based routers. A shared shared auction memory lets agents learn from past bids and refine strategies at test time, so the system keeps improving without retraining. The result: far fewer calls to the biggest model, lower costs, and better final results than using a single large model or standard routing methods.
Not sure where to start?Get personalized recommendations
Data Highlights
153% reduction in reliance on the largest agent across evaluated deep search and coding workloads
235% lower overall execution cost compared to baseline strategies that rely more on the largest model
3Consistent improvement in pass@1 (final-answer accuracy) over the largest agent, achieved with only negligible overhead beyond executing the final chosen plan
What This Means
Engineers building agentic workflows who want to cut cloud costs while keeping or improving solution quality should consider multi-agent coordination strategy auctions. Technical leaders deciding between bigger models or smarter coordination can use this approach to trade model size for system-level gains. Researchers studying multi-agent coordination or agent reliability will find the market-inspired routing and test-time self-improvement ideas directly applicable.
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreYes, But...
Results are shown on deep search and coding benchmarks; other domains (e.g., open-ended dialogue, safety-critical planning) were not evaluated. Sale's performance depends on having a diverse pool of agents whose short plans contain useful, evaluable structure. The auction scoring and memory mechanism add design choices that must be validated for bias, safety, and latency considerations in production settings.
Full Analysis
Smaller language models are cheaper and attractive as building blocks for agentic systems, but their capability does not scale smoothly with task complexity. Performance drops on deep search and coding tasks where long, structured strategies matter. Rather than relying on a single large model or a separate router trained on task descriptions, strategy auctions let each agent submit a short plan for how it would solve the task. Those plans are scored by a transparent cost-versus-value rule that balances expected benefit against execution cost. Winning plans are executed and recorded in a shared auction memory; future auctions reuse and refine successful strategies, giving continual, test-time self-improvement without retraining a separate router. Across the evaluated workloads, this approach cut calls to the largest agent by 53%, reduced overall cost by 35%, and produced higher pass@1 (final-answer accuracy) than the largest agent alone, with only negligible extra cost beyond running the final chosen trace. Compared to routers that pick models from task descriptions, auctions provided better routing and cost savings, showing that system-level, market-inspired coordination can effectively scale small agents for harder tasks. Planning Pattern and Memory approaches help structure and reuse successful strategies, while considering latency considerations for production deployment.
Explore evaluation patternsSee how to apply these findings
Credibility Assessment:
Includes Yoram Bachrach with h-index 41 (top researcher), which strongly elevates credibility despite arXiv venue.