AutoTailor Builds User-Aligned Tool Sets for Web Agents

Summary

AutoTailor is a meta-agentic framework for building and maintaining a compact set of Model Context Protocol (MCP) APIs for web agents. It addresses the problem that automatically discovered browser-automation tools can be large, redundant, and poorly aligned with user demand. Offline, the system converts web trajectories into parameterized browser-automation programs, then uses a Quality Filter to remove APIs with unsuitable granularity or redundant functions and a Usage Likelihood Filter to retain broadly useful capabilities while preserving semantic coverage. On a set of 106 WebArena Postmill tasks, these filters reduced 1,283 unrefined APIs to 87. Its online Dynamic Reselection component monitors task outcomes and API use, adds candidates for recurring coverage gaps, and prunes capabilities that remain unused, producing a 33-API set. With ReAct fallback, the resulting set reached 90.6% correctness versus 87.5% for ReAct alone, while lowering average total request-token cost by 57.8% and latency by 29.4%. Without ReAct, it achieved 60.1% correctness, roughly matching the unrefined set while reducing request-token usage by 94.9%. The results indicate that combining static filtering with adaptation to observed user needs can make web-agent capability management more efficient without sacrificing the reported task performance.