Web agents can employ reusable tools to cut the cost and latency of low‑level browser interactions, yet automatically discovered tool sets are often large, redundant, and poorly aligned with user demand. AutoTailor is a meta‑agentic framework that builds and maintains a compact collection of trajectory‑derived Model Context Protocol (MCP) APIs.
In the offline stage, web trajectories are transformed into parameterized browser‑automation programs. A Quality Filter then discards APIs with unsuitable granularity or redundant functionality, while a Usage Likelihood Filter prioritizes broadly useful capabilities and preserves semantic coverage.
During online operation, the Dynamic Reselection module monitors task outcomes and API usage, identifies recurring coverage gaps, adds relevant candidates, and prunes persistently unused capabilities.
Evaluation on 106 WebArena Postmill tasks shows that offline filtering reduces the initial 1,283 unrefined APIs to 87, and Dynamic Reselection further yields a 33‑API set. With a ReAct fallback, this set achieves 90.6% correctness—higher than ReAct alone’s 87.5%—while cutting average request‑token cost by $57.8\%$ and latency by $29.4\%$. Without ReAct, correctness reaches 60.1%, comparable to the unrefined set, yet request‑token usage drops $94.9\%$.
These findings demonstrate that static filtering produces a compact inventory expected to support core, high‑likelihood tasks, and dynamic reselection tailors that inventory to observed user needs, improving accuracy and latency while sharply reducing token consumption and end‑to‑end cost.
Review