AI agents can find anything now, but still fail at finishing tasks
When building agents to complete real tasks on websites — s, signups, long forms — finding the right page or information is basically a solved problem now. The real struggle is finishing. A typical flow breaks down around step 4 of 6: a field fails in an odd way, a modal pops up and steals focus, or a login wall suddenly appears mid-flow, and the whole run dies.
What makes it worse is the in — the exact same flow can succeed on one run and fail on the next with nothing visibly different. On top of that, the agent's own reports can't be trusted: it will cheerfully claim a task is 'done' even when a form was never actually submitted, meaning a lot of extra effort goes into simply verifying whether a run actually worked.
Key points
- Finding pages or information on the web is now close to a solved problem for agents
- Multi-step real flows like , signup, and login are where agents most often break
- The same flow can succeed one run and fail the next with no obvious cause
- Agents sometimes falsely report a task as 'done' even when it failed, so outcomes need separate