A build-and-break story from production: 32 specialized AI agents assess a live Azure estate against the Well-Architected Framework — and the hardest part was never the prompt. Watch the system produce a plausible finding that is wrong, then catch its own hallucination live. Three reusable patterns for agent output you can actually trust.

LLM agents are confidently wrong, and confident-wrong is fatal the moment an architect ships the fix or an auditor reads the finding. This is a build-and-break story from production. Marc runs a system of 32 specialized AI agents that assess a live Azure and Microsoft 365 estate against the Microsoft Azure Well-Architected Framework — roughly 800 checks per run — and the hardest part was never the prompt. It was making the output trustworthy enough to act on.

This session is mostly demo. On a pre-consented, anonymized Azure tenant you watch the system ground itself in live tenant state via Azure Resource Graph — not the documentation — then produce a plausible finding that is wrong, and then catch its own hallucination, live, as adversarial review passes argue the finding down to nothing with their reasoning on screen. After that, a real finding survives, carrying evidence, a confidence score, and the Bicep or Terraform to remediate it.

Around the demo, the engineering: how the work is decomposed across agents and what that costs in latency and money; the three patterns that actually made agent output trustworthy — real-state grounding, adversarial self-review, and confidence and evidence scoring; and the war stories, including the time it was wrong anyway and the pass that saved us.

You will leave with three reusable patterns for any agentic system you are building, an understanding of why confident-wrong output is a platform risk rather than a UX annoyance, and an honest account of what still breaks.

Bring your skepticism. The demo is built to earn it.