# When your AI hits a dead end: agent feedback

> What happens when Claude or ChatGPT can't do what you asked, where its note goes, and who fixes it.

Last checked 2026-11-06.

## What happens when my AI hits a dead end?

Say you ask Claude to add a photo to step three of a S'mores recipe, and there's no place for one. Before it tells you it can't, Claude sends a short note to the people who run your site manager. Then it tells you plainly what it couldn't do.

The note changes nothing on your site. Nothing is published, edited or undone.

## What does the note say?

- What went wrong: a missing tool, something unexpected, or it got stuck.
- Which site it was working on.
- The tool it used, or the one it wished it had, if there was one.
- What it tried, and what it needed.

It also records who was signed in and which app they used, like Claude or ChatGPT. Each person can send 20 notes a day, so an AI stuck in a loop can't flood anyone.

## Where does it go?

Each note is kept by the site manager, and it also lands on your site's to-do list (we call it the job board). The first note about a problem adds an idea there. An idea is something to talk about. It never becomes a change on its own.

If the same problem comes up again in the same words, the note joins that idea, and the list counts how many times agents have asked.

## How do I read it?

If you own a site, you read your own site's notes, and nothing from anyone else's. Ask Claude: "Summarize the agent feedback on S'mores." It lists the problems asked about most first, with what was tried and what was needed.

Our team sees the notes from every site.

## Who turns it into a fix?

We do. Our team reads every new idea within a day. Each one either becomes a task with a named owner, or is closed with a reason.

Once it's a task, your site manager can start on it.

Whoever does the work, nothing goes live until a person says publish.

## Read next

- [Brain, body, hands and tools: how your AI works on your site](/docs/how-it-works/)