How to find out why an agent team run failed with AI Builder

Learn how to ask AI Builder to read a run and explain why it failed or gave a thin result.

Overview

When a run fails, or its output is wrong, thin or empty, AI Builder can read the run before it answers. It checks the run's status and output, which step errored, called no tools or was sent back to redo its work, and what that step's tools returned. It then explains the cause in plain words and proposes a change to the team. It changes the team only after you agree.

Reading a run changes nothing and costs nothing.

Before you start

• You need to be able to see the run. AI Builder says it cannot find a run that does not exist or that you cannot see.
• A run that is still queued or running is not diagnosed. Wait for it to finish.

Diagnose a run

1. Open AI Builder from the team's page. If you are looking at a run, "the run" means that run.
2. Ask, for example, "Why did the last run fail?" or "Why is the output empty?". With no run named, AI Builder takes the most recent failed run, or the most recent run if none failed.
3. Read the cause. AI Builder names the step and says what it did.
4. For a team that used to work, ask it to compare an earlier good run with the bad one. It points to what differed between the two.
5. Agree to the proposed change, or ask for a different one. AI Builder applies the change to the team and tells you what it changed.

For the card shown on a failed run, see Troubleshooting the "What went wrong" card on a failed agent team run in GTM One.

If AI Builder cannot read the run

• If it says it cannot find the run, check the run id. It must be a run's id, not the team's.
• If it says it could not read the run's details, it answers only from the run's status and output, and does not guess a cause.
• If the team has no runs yet, AI Builder says so and offers to run it. See How to run an agent team from AI Builder.