Every call keeps its transcript and a record of what the agent actually did — so a bad call is evidence, not an anecdote.
From a bad call to the Builder
Open the Calls page and inspect the call in question.
The Builder opens in troubleshoot mode with that call's transcript and diagnostics already loaded. A banner confirms: the diagnostics are in, the fix will be applied to a draft, and callers stay on the live version.
Knowing why it went wrong isn't needed — "it should have offered a morning slot" is enough. The Builder reads the evidence, explains what it finds, and proposes a change.
Success is measured, not guessed
Every agent the Builder designs carries its conversation goals — the outcomes agreed when it was built — in its description. A diagnosis opens with a verdict against them: each goal scored achieved, partial or missed, with evidence from the transcript and the record of what the agent's tools actually did. The call counts as a success only when every must-have goal landed and the agent held to its own prompt.
An agent with no recorded goals is judged against its prompt alone — and the Builder offers to record goals as part of the fix, so the next diagnosis has a yardstick.
Fixes are proposed, not imposed
Troubleshooting is a full editing session. The Builder follows its usual propose-and-confirm rhythm, every accepted fix autosaves into a draft, and the draft can be tested before anything reaches a caller. If the call ran on an older version, the Builder says so — the problem may already be gone.
Resuming a session
While a session is in progress, that team's card on the Agents page shows a Resume troubleshooting button — it reopens the same conversation with the diagnostics still in context.
One call, one session: each hand-off from the Calls page starts a fresh conversation seeded with that call. Chasing a different call means handing it off from Calls again.
When the agent isn't in a team
For a call answered by a standalone agent, the Builder still reads the diagnostics and reports what it finds — but it advises rather than edits, and its suggestions are applied by hand in the agent editor.
Troubleshoot the pattern, not the one-off. If three callers hit the same dead end, hand the clearest example to the Builder and mention the others.