Which timeout is it?
Three different things all look like “the agent timed out”: the model call itself, a tool the agent is waiting on, or the overall run limit. The Deep Dive trace shows which step was running when it stopped — start there.
Model call timing out
- The provider is slow or down — check status, retry later.
- The request is huge — trim the context before retrying.
- The model is too big for the task — a smaller model answers faster; see cost vs capability.
Tool call hanging
A command waiting on input, a server that never responds, or a browser that never finished loading — the trace will show the tool and its arguments. Fix the tool, not the agent: add a timeout to the command, or adjust the task so the tool is optional.
Run limit too short
Long builds legitimately exceed the default run limit. Raise it for that task, or — better — split the task into smaller steps the agent can checkpoint. A task that is half done should resume, not restart: that is the general troubleshooting pattern.