Scaling Small Agents Through Strategy Auctions

Small language models are viewed as a promising, cost-effective approach to agentic AI, yet how their performance scales with task complexity remains unclear. While smaller agents match larger ones on simple tasks, it is unknown when large models become necessary and how to better leverage small agents. In this work, we show that small agents fail to scale with task complexity on deep search and coding tasks, and introduce Strategy Auctions for Workload Efficiency (SALE), a framework inspired by freelancer marketplaces. In SALE, agents bid with strategic plans scored by a cost–value mechanism and refined via shared auction memory, enabling per-task routing and continual self-improvement without training a router. On average, SALE reduces reliance on the largest agent by 52%, lowers overall cost by 35%, and consistently improves pass@1 with only a negligible token overhead. In contrast, established routers either underperform the largest agent or fail to reduce cost. These results suggest that small agents can be effectively “scaled up” through coordinated allocation and test-time self-improvement. More broadly, they motivate a systems-level view of agentic AI in which gains come less from ever-larger individual models and more from market-inspired coordination mechanisms that organize heterogeneous agents into efficient, adaptive ecosystems.

Smaller AI agents are attractive for their low running costs, but it remains unclear when they suffice and when the task genuinely demands a larger, more expensive model. We set out to understand where small agents fail, and whether there is a more discerning strategy than simply reaching for the biggest model available. Across deep-search and coding tasks, we found that small agents cope well with easy problems but fall behind as complexity rises. Drawing on the logic of freelancer marketplaces, we built a system called SALE in which agents compete for each task by submitting a plan describing how they would solve it. The plan functions as a bid: a central mechanism appraises its cost against its likely value and awards the work accordingly. A shared memory of earlier auctions lets the system sharpen its judgments over time, with no retraining and no separate model dedicated to assigning tasks. In our experiments, SALE calls on the largest agent less than half as often, cuts total cost by more than a third, and answers more accurately. Competing approaches to routing, by contrast, tend to cost more, perform worse, or both.