If your agent suddenly seems stupid: openclaw tools.profile-Fail

Florian · · ~4 Min. Lesezeit
AI Blog
If your agent suddenly seems stupid: openclaw tools.profile-Fail

Last updated: March 16, 2026

You switch from the top model to a cheaper model (or slip into a fallback) and suddenly your agent seems thwarted: a lot of explanation, little execution. This feels like “model got dumber”, but it is often a different error: Capability drift due to an incorrect tool profile.

We had exactly this practical case in the OpenClaw setup: With tools.profile: messaging only messaging/sessions were available - but no exec, no file system, no browser automation. Result: The agent plans properly, but can no longer carry out its own steps.

In this guide you will get: quick diagnosis, clear fixes, break glass recovery and concrete tips on when it is better to use a strong model like Codex or Claude Code for troubleshooting.

Why this happens: Capability drift instead of an IQ problem

A model change not only changes the quality of the response, but sometimes also indirectly changes the available tool scope. If the profile falls to ‘messaging’, the agent loses operational capabilities. This is not a “prompting problem”, but a runtime/policy issue.

  • Symptom: Agent explains steps to you instead of executing them.
  • Misdiagnosis: “The model is too weak.”
  • Real cause: Tool rights are missing (exec/FS/browser gone).

The typical warning signals in practice

  • The agent repeats the same plan in multiple responses.
  • It keeps asking for manual shell steps even though the task could be automated.
  • File changes are not made (read/write/edit is actually not used).
  • Browser/web steps are announced but not executed.
  • Error texts sound like authorization/tool ​​limits instead of technical errors.

Quick diagnosis in 2 minutes

Proceed strictly in this order:

  1. Check session status: Which model is active? Was there a fallback?
  2. Check tool profile: Is tools.profile unexpectedly set to messaging?
  3. Mini Probe: Is it possible to run a harmless exec command?
  4. File Sample: Can the agent read/write a test file?
  5. Browser Sample: Can a snapshot/navigation be done?

If points 3-5 fail, it is almost certainly not a “stupid model”, but rather a lack of capabilities.

The quick fix (proven)

  • Set tools.profile to full.
  • Start session/task fresh (do not continue to use a half-broken state).
  • Run quick smoke checks (see below).

That was exactly the solution for us: After resetting to ‘full’, the entire tool chain came back (exec, file system, browser, memory), and the loop was immediately gone.

Break-Glass: When the agent can no longer save himself

  1. Stop open hanging tasks.
  2. Correct profile/config explicitly.
  3. Start a new clean session (without legacy context).
  4. Just run a small sample task.
  5. Only then reactivate productive jobs.

Pro tip: For critical automations, it is better to do a quick “hard reset” rather than continue debugging in a partially defective state. This usually saves hours.

If OpenClaw hangs: Use a strong LLM as a co-debugger for troubleshooting

Important in practice: If OpenClaw is already stuck in a weak fallback, “ask OpenClaw for the solution” often only helps to a limited extent. In precisely this condition, tool calling is often limited or unreliable. You then need an external, strong copilot for the diagnosis.

Proven shortcut: Use Codex, Claude Code or a comparably strong LLM for the analysis. Have a structured recovery plan built there and then import it specifically into OpenClaw.

  • Step 1: Briefly summarize symptoms + logs + current model/fallback.
  • Step 2: Have the strong LLM create a prioritized hypothesis list (config, tool profile, rights, session state).
  • Step 3: Have concrete test commands/checks generated (no generic blah blah).
  • Step 4: Roll out fixes in small order and verify after each step.
  • Step 5: Only then switch back to normal operation.

This often saves massive amounts of time because you’re not trying to use an already limited agent to debug its own limitation.

Conclusion

If your agent suddenly seems “stupid”, check capabilities first instead of just model quality. In our case, the core error was an incorrect tool profile — and the fix to tools.profile: full immediately normalized operations. With short smoke tests and clear UAT criteria, you can prevent such an error from slipping back into the live flow unnoticed.

CTA: For critical incidents, first use a strong external co-debugger (e.g. Codex or Claude Code), create a clear recovery plan and then transfer the fixes to OpenClaw in a structured manner. This usually saves you the most time, especially in weak fallback states.

Weiterführende Artikel