Browser agent failure mode: the page said no, but the agent thought it succeeded
ahstanin · reddit · 2026-09-03
A developer building browser tooling for agents describes a subtle, recurring failure mode: when a web form rejects a submit, the page usually prints a message near the fields instead of adding new controls. If an agent's action result only reports structural change, a refused submit looks identical to a successful one — the agent proceeds, every later action runs against a screen that never advanced, and the task fails steps later somewhere unrelated. Screenshot-based agents have it harder, since the refusal is just a few red pixels the model must notice.
His fixes:
- Make the action result carry what the page actually said, not just structural changes;
- Treat refusal as a first-class outcome, distinct from success and error, and use it as a stop condition for the rest of a batch;
- Sample visible text in observations, not just the control tree, or validation messages get missed.
The key argument: refusals need a different recovery path and can't be lumped in with success or errors.
More from coding & agent
- mcp-doctor: a static analysis linter for MCP servers that surfaced real bugs in 12 popular repos — Wide_Imagination_970 · 2026-09-03
- Stripe launches Directory preview to help AI agents find and use external providers — jeff_weinstein · 2026-09-03
- Akkru launches financial data MCP so AI agents can trace numbers back to filings — camiggggg · 2026-09-03
- Matt Shumer explains Fable's architecture: locked world structure lets versions share one world — mattshumer_ · 2026-09-03
- GLM-5.3-Flash priced at 0.06x: Factory confirms all-day droid usage without rate limits — matanSF · 2026-09-03
- Qwen 3.8 users report better results forcing reasoning effort to low, llama.cpp flag shared — Jorlen · 2026-09-03