Your AI is going in circles. Here is how to get it unstuck

Sam asks the agent to make a spreadsheet’s totals match the bank statement. It changes a formula and says it is fixed. The totals still differ. Sam asks again. It changes a different formula, then changes the first one back. Sam asks again, more firmly this time, and gets a longer apology and the same result.

If that sounds familiar, you are not doing anything wrong. Every model gets stuck sometimes. What matters is noticing early and changing something, instead of asking the same thing again.

Spot the loop

Here are the signs an agent is going in circles:

• The same fix is tried a third time.
• It says “fixed” or “done” and nothing has changed.
• It undoes something it did earlier, then does it again.
• Its answers get longer while saying less.
• It keeps apologising and promising to be more careful.

A good rule: if the second attempt at the same thing fails, change your approach before the third.

Why asking more firmly does not work

When you repeat yourself in capitals, the agent still sees the problem the same way. It still holds the same wrong belief about it. And the chat is now full of its failed attempts, which pull it back towards the same ideas. More pressure gets you the same answer, said with more apology.

So change something. There are five good moves, and they work in almost any AI tool.

1. Switch the model, or the family

The quickest way out of a loop is often a different model. Each model has its own habits and blind spots, and one that keeps missing something may be stuck in a way another is not.

Models from different families, such as Claude (Anthropic) and GPT (OpenAI), are trained very differently, so they notice different things. If you were on Claude, try GPT. If you were on GPT, try Claude.

ChatGPT only offers OpenAI’s models and Claude only offers Anthropic’s. In Interweave OS you pick another model in the same chat and send your next message, and the new model is handed the whole conversation, so you do not explain it all again.

Either way, ask for a fresh read, not just another attempt.

Try saying:

We’ve been going in circles. Read the conversation, tell me what you think is really going on, and try a different approach from the ones above.

Asking what is really going on matters. It stops the new model simply carrying on with the last failed idea.

Write your next message for the family you moved to. GPT is literal, so spell out the goal, every requirement and what to do when it is finished. Claude fills gaps, so say what success looks like and what it must not change.

Or move up

Sometimes the problem is just hard. Pick a bigger model, or more thinking effort if the model offers it. More effort is slower and uses more, but helps with tricky reasoning. Once the hard part is solved, come back down for the routine work.

Or bring in a helper

If your tool supports helpers, you can keep your current model and ask it to bring in one from another family just for the stuck part.

Try saying:

Get a helper from another family, such as GPT or Gemini, to look at this problem on its own, without seeing your attempts, and suggest a fix. Then try its idea.

2. Start fresh with a handover note

A very long chat can bury the point under old attempts. The agent has to keep all of it in mind, including every idea that did not work, and it may keep drifting back to them. Sometimes the best move is a clean start that keeps what you learned.

Start fresh when:

• The chat has gone round in circles and switching model has not helped.
• The chat has drifted onto a new topic. A new topic deserves a new chat.
• The agent seems to have forgotten things you agreed much earlier.

Here is how:

1. Ask the agent in the old chat for a handover note.
2. Read it and fix anything wrong. This is your chance to drop the bad ideas.
3. Start a new chat and paste the note as your first message, with what you want next.

Try saying:

Write a handover note: the goal, what we decided, what we tried that did not work and why, open questions, and where the files are. I’ll start a new chat with it.

Step two is the one people skip. Do not. A handover note that still contains the wrong belief will carry the loop straight into the new chat.

3. Find the wrong assumption

Many loops come from one wrong belief early on. The agent assumed something, built on it, and every fix since has been built on the same mistake. No amount of fixing the top layer helps while the bottom is wrong.

So ask it to show you what it is standing on.

Try saying:

Stop for a moment. List everything you are assuming about this problem: about the data, the files, what I want, and how it should work. Mark the ones you have not checked.

Read the list. One of the unchecked ones is usually the problem. In Sam’s spreadsheet, the agent had assumed every amount included GST. Half did not. Once Sam said so, the totals matched in one step.

4. Shrink the problem

When a job is big, a failure anywhere makes the whole thing look broken, and it is hard to see which part is wrong. Make it small. Get one piece working, prove it, then move to the next.

Try saying:

Forget the full report for now. Just get the January figures right and show me how you checked them against the bank statement. Then we’ll do February.

Small pieces also make progress visible. You can see what works, and the agent has less room to wander.

5. Give it what it cannot see

The agent knows only what is in the chat, its instructions, and the files it can read. If it keeps getting something wrong, ask yourself what you know that it does not. A screenshot of what you see, the exact wording of an error, or a sample of the right answer often unlocks it straight away.

• Not: “It’s still wrong.”
• Better: “The total for March should be $48,210. Yours is $52,900. Here’s the bank statement for March.”

The second message turns a guessing game into a checking job.

When usage runs out or a model is slow

A cousin of the loop is the stall: the model you were using hits its usage limit halfway through, or is slow today. Neither is a reason to stop. Switch to another family and keep going. If you switch partway through a job, ask the new model to check what was finished first, because the last one may have stopped halfway through a step.

Try saying:

The other model ran out of usage partway through. Pick up from where it stopped: check what was finished, then carry on with the rest.

Being comfortable with both families is useful in itself. If you only ever use one, a limit or an outage stops your day.

Stop the same loop happening again

Once you find what caused a loop, ask whether it could happen again. If the agent got stuck because it did not know that half your amounts exclude GST, the next chat will not know either. Put the fact in your standing instructions, or your team’s, so every new chat starts out knowing it. Interweave OS can also help here: tell Improve “We’re going in circles” and it works out why the chat went wrong and suggests what to change in your request or your instructions.

How Interweave OS does it

Interweave OS is built so a stuck chat never means starting over. Switch to another model or family in the same chat and it carries on with the whole conversation intact. Ask for a helper from another family to look at just the stuck part. And while the agent works, a quiet line under the messages tells you when it looks stuck, such as retrying the same failing step, so you can steer it a new way early.