A browser agent failure mode that's easy to miss: the page says no and the agent keeps going
ahstanin · reddit · 2026-08-28
A recurring failure mode when building browser tools for agents: when a form rejects a submit, it usually adds no new controls — just a message near the fields. If the action result only reports structural change, a refused submit and a successful one look identical, so the agent proceeds and every remaining action runs against a page that never advanced; the task fails three steps later somewhere unrelated. Screenshot-based agents have it worse: the refusal is a few red pixels the model must notice and interpret.
The fix: make the action result carry what the page said, not just structural changes, and treat a refusal as a stop condition for the rest of the batch. Two general suggestions:
- Sample visible text in your observation, not just the control tree, or you'll miss every validation message.
- Make refusal a first-class outcome distinct from success and error — recovery means fixing the named fields, not retrying the click.
Notably, the author's own test suite missed this: tests re-read the page after each step and saw the error, but the per-action result was blind — exactly what a batching agent reads.
More from coding & agent
- AI documents long-term tasks to avoid wasted tokens — draginol · 2026-08-28
- AI Agent uses task comments to document long-term projects — draginol · 2026-08-28
- Implementing Automatic Model Routing Based on Intent — chadwell · 2026-08-28
- Google's Harness Engineering: Agent = Model + Harness Replaces Prompt Engineering — anselm · 2026-08-28
- GLM-5.3 Released: 50% Coding Gain, Emergent Cyber Capabilities — jacek2023 · 2026-08-28
- zai releases GLM-5.3 open-weight model for agentic coding and defense — zai-org · 2026-08-28