When Your AI Agent Keeps Going in Circles: What to Do

You ask your AI coding agent to solve a problem. It tries something. The error stays. It tries something slightly different, and the error stays. Twenty minutes later you have a pile of changes, the same bug, and a growing suspicion that the agent is not going to get there on its own.

This happens with Claude Code, Cursor, Codex and every other agent. It is not a sign that you are doing something wrong. Here is how to recognise it, what to try before you give up, and when a person is simply faster.

How to tell your agent is going in circles

• Same error, new explanation. The message on screen has not changed, but the agent keeps announcing a different cause.
• It undoes its own work. A change from ten minutes ago quietly disappears, then comes back.
• It says it works, but it does not. The agent reports success while the page, the test or the deploy still fails.
• The changes keep growing. A small bug now touches files you never mentioned.

Three rounds of this is enough. Past that point, more prompting rarely helps, and every extra round makes the project harder to untangle.

Three things to try yourself first

1. Stop and save the current state. Commit or copy your project before the next attempt, so you can always go back. If something already changed that you did not ask for, our guides on Claude Code, Cursor and Codex explain how to undo it.
2. Start a fresh conversation with a short, exact description. Give the agent the error message word for word, what you expected, and what happened instead. A long chat full of failed attempts often sends the agent in the wrong direction.
3. Ask for one cause before any change. Tell the agent to explain why it thinks the error happens and to change nothing until you agree. Wrong guesses are cheap to read and expensive to apply.

When a person is faster

If the same problem survives a fresh start, a person usually sees in minutes what the agent has missed for an hour: a setting in the wrong place, a mismatch between two parts of the app, a change that needs a decision rather than more code. This is also the right moment for a second pair of eyes before launch.

On humans fix ai you describe what you need in plain words, and a developer we have approved by hand looks at it with you on a free video call. Most problems are solved right there. If more work is needed, the developer names one fixed price, you can say no, and you pay only after you approve the result. More on how this works is on our human in the loop page.

Let your agent ask for you

If you use an agent that supports MCP, you can connect humans fix ai once. The next time the agent is going in circles, it can offer to ask a developer. It writes the request, shows you the exact text, and sends nothing until you say yes. Setup takes about a minute: see how to connect it.

The short version

Three failed attempts means it is time to change the approach. Save your work, restart with an exact description, and ask for a cause before a change. If that does not work either, ask a developer. The first call is free.