Falsifiable Commitment Planning for Self-Correcting Web Agents
2026-07-27 • Artificial Intelligence
Artificial Intelligence
AI summaryⓘ
The authors developed FCPAgent, a system designed to help web automation tasks succeed over long sequences of steps. Unlike previous agents, their approach tracks clear evidence to verify if each planned action still makes sense as it goes along. If a problem is detected, the system can identify where the issue happened and fix only that part to get back on track. Their method showed better success rates on complex tasks compared to earlier methods.
web agentslong-horizon planningfalsifiable commitmentsevidence verificationplan-test-repair loopskill reuseLLM diagnostic verificationscope-aware repairWebArena benchmark
Authors
Guangyi Liu, Huan Zhao, Quanming Yao
Abstract
Long-horizon web agents often go off track before final failure: a trajectory can remain locally plausible even after the current state, reused skill, or plan assumption no longer supports the user instruction. Existing agents can plan, reflect, or reuse experience, but their plans rarely specify the evidence under which an active step should still be trusted. We propose FCPAgent, a falsifiable commitment planning framework for robust long-horizon web agents. FCPAgent represents each plan step as a Falsifiable Commitment Unit (FCU): a subgoal grounded in a reusable skill, together with confirming evidence, falsifying evidence, and a confidence score. Execution is organized as a plan-test-repair loop. The hybrid commitment testing module checks candidate actions before they modify the browser and checks observations after execution; for efficiency, it combines lightweight evidence matching with LLM-based diagnostic verification. When evidence falsifies a commitment, scope-aware repair localizes the contradiction to the execution, skill, or planning level and revises the smallest adequate part. On WebArena, FCPAgent achieves a 13.8% relative improvement in average success over the strongest baseline, with especially large gains on long-horizon tasks.