FCPAgent: Falsifiable Commitment Planning for Self-Correcting Web Agents
The newly introduced FCPAgent framework tackles the issue of long-horizon web agents that deviate from their intended path prior to ultimate failure. While current agents can plan, reflect, or leverage past experiences, they seldom define the conditions under which an ongoing action remains reliable. FCPAgent conceptualizes each step in a plan as a Falsifiable Commitment Unit (FCU), which encompasses a subgoal tied to a reusable skill, supporting and contradicting evidence, along with a confidence score. Its execution follows a plan-test-repair cycle. The hybrid commitment testing module evaluates potential actions before they alter the browser and assesses observations post-execution, merging lightweight evidence matching with LLM-based diagnostic checks for enhanced efficiency. This framework is detailed in a paper on arXiv (2607.24167).
Key facts
- FCPAgent is a falsifiable commitment planning framework for robust long-horizon web agents.
- Each plan step is represented as a Falsifiable Commitment Unit (FCU).
- FCU includes subgoal, reusable skill, confirming evidence, falsifying evidence, and confidence score.
- Execution follows a plan-test-repair loop.
- Hybrid commitment testing module checks actions before and after execution.
- Combines lightweight evidence matching with LLM-based diagnostic verification.
- Paper published on arXiv with ID 2607.24167.
Entities
Institutions
- arXiv