AI/ ai · web-agents · benchmarks · research

Web Agent Benchmarks Miss Most of What Goes Wrong

WebStep tracks every intermediate step of a web agent's session, revealing that agents with identical success rates fail in completely different ways.

A new benchmark finds that web agents with nearly identical task-completion rates fail in entirely different ways, a gap that current evaluations miss entirely.

Researchers built WebStep, a collection of 1,800 task instances that pairs a standard browser interface with a semantic state tracker running silently in the background. Instead of only recording whether the agent finished the job, WebStep logs every intermediate state and transition, essentially a flight recorder for AI browsing sessions. That process data revealed something pass/fail scores obscure: three leading web agents all completed tasks at roughly the same 31-33% rate, yet their failure profiles were almost opposite. On a housing-search task set, OpenAI's CUA outperformed Qwen3.5 by 23.7 percentage points on committing actions but trailed by 15.6 points on filtering, the kind of specific diagnostic a binary outcome score cannot surface.

Most of the web-agent field is currently judged on benchmarks where you either complete the task or you don't. That's useful for ranking models on a leaderboard but tells a development team almost nothing about what to actually fix. WebStep's process-level approach lets a researcher say "this agent executes well but explores poorly" rather than "this agent scores 32%", and the gap widens sharply as tasks get harder.

Whether the research community adopts process-level evaluation or keeps chasing terminal-success leaderboard numbers will say a lot about whether the field is serious about improving agents or just scoring them.

TR

The Revision

Written by an AI system from the public sources credited above. How we write →