An agent seems stuck
A run hasn't moved in a while. Check for an ask, watch the terminal, then pause, take over, hand off or restart the run.
On this page
Most “stuck” runs are actually waiting for something: a person, a limit, or a provider. Work through the causes below in order; the first two cover most cases.
Symptoms
- The run’s step hasn’t changed for longer than usual.
- The run shows Waiting but you don’t see why.
- The conversation has gone quiet and no new files appear under Changes.
Likely causes and fixes
It’s waiting for someone
A run that needs a person stops and puts an ask in the Inbox: approve a plan, allow a command, answer questions, or review a pull request. The ask may be in someone else’s Inbox, such as the issue owner or a CODEOWNER.
- Open the run and look at the top of the conversation for an ask card.
- If it’s yours, answer it with the button on the card, such as Approve plan or Allow once.
- If it’s someone else’s, use Hand off to send it to a teammate who can answer, or ask them directly.
See Approve a plan and Answer an agent’s questions.
It’s working, just slowly
Long builds, test suites and large refactors can run quietly for a while. Open the Terminal tab to see live output. If commands are still running, the agent is fine. For more detail, open Show in Fleet from the run menu to watch every terminal at once.
It’s going in circles
If the terminal shows the same command or error again and again:
- Choose Pause after this step from the run menu. Nothing else runs until you say go.
- Tell the agent what to change in the conversation, then choose Resume run.
If it still loops, take control yourself with Take over in terminal (or press T), fix the problem, and hand control back. You can also choose Hand to another agent… to let a different agent, for example tester for a flaky test, pick it up.
A limit was reached
Runs pause and ask when they reach a limit; they’re never stopped silently.
- Budget: the run shows Paused · budget. See A budget paused my run.
- Timeout: each workflow has a timeout, and steps can have their own. A run that hits one stops with a message saying which. Raise it in the workflow’s Limits (see Limits and budgets), then restart.
The runtime or model provider stopped answering
- A runtime subscription ran out. Claude Code, Codex and Gemini CLI plans have their own limits. In Settings → Models, When a runtime subscription runs out decides whether runs continue on Duun credits or pause. If it’s set to pause, switch it, or wait for the subscription to reset.
- The provider is rate-limiting you. Set a Fallback model in Settings → Models. It’s used when the primary provider is rate-limited.
- Your own key stopped working. If you use your own keys, check the provider in Settings → Models with Test connection. See Models and BYOK.
The machine went away
If the run’s sandbox or computer stopped, the run waits. Open the machine from the run header. For SSH computers, check the host is up and still trusts Duun’s key.
Start over cleanly
If none of this helps, choose Restart run from the run menu. If you’d rather stop, Cancel run… stops the agent after its current tool call and keeps the worktree and thread, so you can restart later. See Stop, pause or retry a run.
Still stuck? Contact support with the run id (for example PLT-482) and roughly when it stopped moving.