The LeadMgr build needed 11 rounds of adversarial QA to surface the failure modes hiding behind "passes locally on a clean address."
Your agent team exists. It's not working.
The failures come in shapes. Prompts that produce the wrong output. Agents that step on each other's context. Token costs that spike without a corresponding spike in useful work. Output quality that degrades over time. Infrastructure that breaks at 3 AM because someone changed an upstream API.
Each failure mode has a different root cause. The wrong fix makes it worse.
The Diagnostic Pipeline. Walk the pipeline. Find the failure. Root cause it. Fix it. Verify the fix holds.
The Coach traces your pipeline from brief to output. Every agent call, every handoff, every decision point. This is the equivalent of reading the code, except the "code" is a series of prompts, models, and orchestration logic.
The Coach identifies the failure modes. Not just the symptom you reported. The ones you haven't found yet. LeadMgr's email verification passed locally on clean addresses. Eleven adversarial passes found the modes hiding behind that.
You get a root cause document. What broke, why it broke, what the fix costs. No jargon. No recommendations you can't act on.
The factory applies the fix. Prompt rewrites, pipeline rearchitecture, model swaps, whatever the root cause demands. The fix goes through the same build pipeline as a new build.
The factory re-runs the pipeline. The failure mode is gone. The fix doesn't introduce new failures. The build log gets an entry.
The Coach assigned: amelia. amelia shipped. The QA agent found what amelia missed. That's the system working.
"Tell us what you need fixed."
If your agents produce output that looks right but isn't, you need debugging. If your token costs are climbing without a corresponding improvement in output, you need debugging. If something broke and you don't know why, you need debugging.
The Coach traces your pipeline from brief to output. Every agent call, every handoff. You get a root cause document with the fix and its cost. Takes a few hours.
Depends on the failure mode. Prompt fixes land in hours. Pipeline rearchitecture takes days. The audit tells you which one you need.
The verification step tests the fix against the failure mode. If the same failure recurs, we re-audit at no charge. The build log gets an entry either way.
Yes. The pipeline doesn't care who wrote it. We trace it, find the failure, fix it. The build log documents every step.