New Web Agent Framework Boosts Long-Horizon Task Success by 13.8%

Web agents built on large language models tend to fall apart the longer a task drags on. A booking flow that requires ten sequential clicks gives the model ten chances to quietly commit to a wrong assumption, and by the time the error surfaces, the whole plan has already drifted off course.

A new paper introduces FCPAgent, a planning framework designed to catch that drift before it compounds. Rather than letting an agent execute a plan and hope each step holds up, FCPAgent breaks the plan into what its authors call Falsifiable Commitment Units (FCUs), discrete assumptions that can be actively tested against evidence as the task unfolds.

The core mechanism is a hybrid commitment testing module. It pairs a lightweight evidence-matching check, useful for catching obvious mismatches cheaply, with a more expensive LLM-based diagnostic verification step reserved for ambiguous cases. That combination lets the agent flag a faulty commitment early rather than discovering it several steps downstream.

Where the gains show up

Tested on WebArena, a standard benchmark for evaluating autonomous web-browsing agents, FCPAgent posted a 13.8% relative improvement in average success rate over the strongest baseline in the comparison.

The improvement was not evenly distributed across task types. The paper reports that FCPAgent's advantage widens specifically on long-horizon tasks, the multi-step workflows where baseline agents historically struggle most, which is consistent with the framework's core premise: catching bad commitments matters more the longer an agent has to act on faulty premises before anyone notices.

Why the distinction matters

Most agent-reliability work to date has focused on better planning or richer tool use. FCPAgent instead targets verification, treating each planning step as a testable claim rather than an assumption to be executed on faith.

That framing has practical implications for anyone building production web agents, where a single unchecked misstep in a long workflow (an incorrect form field, a misread page state) can silently derail an entire session. Whether the falsifiable-commitment approach generalizes beyond WebArena to messier, real-world sites remains an open question the paper does not address.

The work was submitted to arXiv on July 27, 2026, and the authors have not indicated whether the code or evaluation harness will be released publicly.