AT-AT: Automatic Tool Optimization from Agent Traces

Efforts to improve LLM-based agents typically focus on modifying the agent itself, yet success also depends on the environment in which an agent acts. Tools mediate this interaction by defining the actions available to agents and the feedback they receive during execution, making the tool environment a complementary target for optimization. We introduce AT-AT (Automatic Tool Optimization from Agent Traces), an agentic tool optimizer consisting of a lightweight coding agent that reads agent trajectories from prior agent executions, as well as agent evaluation results, and optimizes tool descriptions and implementations accordingly. On $\tau^2$-bench-verified, AT-AT achieves higher task completion rates at lower cost than strong prompt and tool optimizer baselines, with paired statistical tests confirming improvements despite evaluation stochasticity. We further demonstrate that AT-AT generalizes across domains by improving performance on the coding agent benchmark TBLite. In an enterprise production setting, applying AT-AT to production logs substantially reduced MCP tool call error rates. By turning accumulated execution traces into updates to shared tool environments, AT-AT provides a path toward continual improvement of agent systems through environment-side learning.