Skip to main content

An agent is stuck

For operators who see an agent at work with no progress for several minutes: identify the cause and get the work moving again.

Quick checks (in order of likelihood)​

1. The agent is waiting for your answer​

By far the most frequent case. The agent tried a sensitive action (writing a file, sending an email, calling an external tool) and is waiting for your consent, or it asked you a question.

Solution:

  1. In the sidebar, click To review.
  2. Everything that waits for you is in one list: consents, questions, paused tasks and batch items. The Decisions filter keeps the consents only, and each item says where it comes from.
  3. Answer on the card: Once, Always for this agent or Refuse. The conversation, the task or the batch concerned resumes at once. See Approve or refuse an agent action for the details.

2. The agent is waiting for an answer from an external tool​

A call to an MCP server (Notion, GitHub, Slack) or to a remote API can take time if the service is slow or unreachable.

Solution:

  1. In the sidebar, open Agents.
  2. Find the agent concerned in the list and right-click it.
  3. Choose Logs. A sliding panel opens from the right with the history of its tasks.
  4. Use the search field at the top of the panel to type a tool name (notion, github, fetch, and so on): only the tasks whose input or result mentions that tool stay visible.
  5. If the last running task has been frozen on that tool for several minutes, open Connections in the sidebar and test the matching connection.

3. A credential or an authorization expired​

OAuth tokens (Google, Notion, GitHub) expire regularly. The agent then fails on every call to that service.

Solution:

  1. Open the Logs panel of the agent (see check 2 above).
  2. Filter on the FAILED status. Type 401, 403, expired or unauthorized in the search field to target authorization errors.
  3. If you find some, go to Connections and reconnect the service concerned.

4. The agent loops on the same action​

Some agents can get stuck on a step they retry again and again. Apollia applies a limit on the number of steps, whose ceiling depends on the autonomy tier: it is lower in assisted and higher in bounded_autonomous or long_autonomous.

Solution:

  1. In the Logs panel of the agent, look for several consecutive tasks with the same input or result.

  2. If the work runs in a conversation, open that conversation and click Stop in the message box.

  3. Ask again once you have fixed the cause (instructions too vague, missing tool).

    Note: an agent has no start, stop or on/off control of its own: it is loaded by its first task and released once idle. If a task that does not run in a conversation never ends, quit Apollia and open it again.

5a. The agent takes longer than expected in the supervised or bounded_autonomous tier​

In the supervised, bounded_autonomous or long_autonomous tier, the agent checks its own work after execution and attempts a self-correction if needed. This lengthens the apparent duration before the agent declares itself done. It is normal behaviour, not a blockage.

If the duration looks excessive to you, open the Logs panel (see View an agent's logs). Verification has no status of its own: the task stays WORKING while the agent checks itself. What tells you the loop is failing to converge is a task that stays working far past the point where its result should have landed. Stop the work, then rephrase the instructions.

5b. A dependency is missing (tool, file, model)​

The agent may require a connection that is not set up, a file that cannot be found or a local model that has not been downloaded.

Solution:

  1. In the Logs panel, filter on FAILED and type not found, missing or unavailable in the search field.
  2. Add the missing connection from Connections, or download the model from the AI screen with Discover models.
  3. Ask the agent again.

If nothing works​

  1. Stop the work in its conversation, then ask again with a shorter, more precise request.
  2. If the blockage comes back immediately, open the agent in Agents and read its sheet again: its mission, its model, its accesses and its sources.
  3. Collect the logs for support: open the Logs panel of the agent and click the Copy logs icon in its header, next to the Refresh icon (hover an icon to read its name). The tasks currently displayed (filters and search taken into account) are copied to the clipboard as text, ready to paste into a report. A message confirms the number of tasks copied.

Technical reference: Apollia reference.