ChatGPT Agent Not Working: Diagnose and Resume the Task

When ChatGPT Work (the autonomous task agent in ChatGPT) stops or refuses to run, the first step is to narrow the cause before retrying. An identical symptom, the task stalling or the agent failing to start, can come from four different sources: a service outage, an account or plan limitation, a permission block the agent hit mid-task, or an error in the task execution itself.

Distinguishing these quickly saves time. Retrying a task against a live outage achieves nothing. Retrying with the wrong permissions set hits the same wall. A tool like fenbs helps by keeping a record of what the agent completed before it stopped, so a retry starts from the right point rather than from scratch.

Quick diagnosis steps

1. Check the OpenAI status page at https://status.openai.com. If ChatGPT or its API shows an incident, wait for it to resolve before retrying.
2. Refresh the page or restart the ChatGPT app. A stale browser session or a connectivity blip often causes an apparent failure.
3. Confirm that your plan includes ChatGPT Work. Agent features are plan-gated; if you recently changed plans, the feature availability may have changed.
4. Review the conversation for a mid-task stop message. ChatGPT Work often explains why it stopped: a permission it needed, a tool that returned an error, or a decision it could not make.
5. If the task involved a connected integration (such as a Google or Microsoft service), check that the connection is still authorized.
6. Before retrying, note what the agent completed. If it finished three of five steps, a retry that repeats the first three wastes time and may produce duplicate outputs.

Service outages

OpenAI's status page at https://status.openai.com shows live incident reports for ChatGPT and the underlying API. If there is an active incident, the agent may fail to start, return errors partway through a task, or behave inconsistently. Wait for the incident to be marked resolved before concluding there is a configuration problem.

Partial outages also occur: the chat interface may be working while a specific tool the agent relies on (such as web search or code execution) is degraded. If the agent fails only on tasks that involve one kind of tool, check whether that service is listed separately on the status page.

Account and plan issues

ChatGPT Work is available on plans that include agent-level access. If you see the agent option in the interface but tasks fail immediately, verify your plan on OpenAI's account page. Usage limits can also cause tasks to stop; some plans cap how many agent tasks run per day or per month.

If you recently changed plans or renewed a subscription, sign out and back in to ensure the session has the updated entitlements. A cached session sometimes holds the old plan state.

Permission stops mid-task

ChatGPT Work stops and asks for confirmation when it reaches an action it cannot take without human approval: sending a message on your behalf, modifying a file, or making a change in a connected service. If you do not respond, the task remains paused. This is intentional; the agent is designed to wait rather than proceed without consent.

Check the conversation thread for a waiting prompt. If the agent is asking a question, answer it and the task should continue. If a specific permission was revoked in a connected integration, re-authorize the connection and retry from the point of failure.

For connector-specific failures, ChatGPT connectors not working covers how to diagnose and re-authorize specific integrations.

Task execution errors

A task can fail because the goal was ambiguous, the required data was unavailable, or a tool the agent called returned an error. In these cases the agent usually stops and reports what it tried and what went wrong. Read the explanation before retrying.

Common fixes: narrow the task scope, provide missing data directly in the prompt rather than asking the agent to fetch it, or split a large task into smaller steps. A task that asks the agent to "research and write and format and send" in one go is more likely to hit an error than the same work broken into separate tasks.

If your task involves coding steps, cursor agent not working and Claude Code not working cover similar diagnosis patterns for other agent-enabled tools you might compare.

Resuming after a fix

Before retrying, check the conversation thread and identify which steps completed successfully. If the agent drafted a document before stopping, you do not need to regenerate it; you need only to continue from the next step.

An outside record of what the agent completed makes this easier. Note the finished steps, the useful output and the remaining work in a task note or a project tool. Then start a fresh conversation with that summary and ask the agent to continue from the identified point. This also works when you want to hand the remainder to a different AI tool. See the overview of ChatGPT agent mode for how multi-step tasks are structured.

How to do this with fenbs

1. Create a free account and add a project for the work the agent is handling.
2. Write the task goal and its steps as a fenbs task before you give it to the agent. This becomes the source of truth for what is expected.
3. When the agent stops, update the task with what it completed: note the steps that ran, any output worth saving and the next action.
4. Connect a supported assistant from Integrations and ask it to read the task before retrying. It picks up from the recorded progress rather than starting over.
5. After the task finishes, save a progress note with the result and what the next session should do. If the same error recurs, record it as a lesson to check before similar work.

fenbs does not run or retry agent tasks for you. It keeps the task record and progress notes that make a retry or a handoff to a different tool faster and more accurate.