Your Evals Passed. Your Agent Just Emptied a Database.

Green CI. Clean eval suite. Staged rollout looked fine.

Then your agent hit an edge case at 2am, chained three tool calls nobody anticipated, and did something irreversible. Your evals didn’t catch it. They couldn’t. They were never designed to.

This is the talk the agent ecosystem needs to have right now — because we’ve spent two years obsessing over eval quality and almost zero time on what happens after the agent is deployed and acting in the world.

Evals are a pre-flight checklist. Your agent is already in the air.

I’ll show you exactly where the gap lives — and how to close it with a runtime verification layer you can bolt onto any agent stack this week:

• Why the gap is structural, not a skill issue. Evals test known distributions. Agents encounter unknown ones. No eval suite survives contact with real users at scale.

• The failure taxonomy nobody publishes. Tool call hallucination under load, trust decay in multi-step chains, silent state corruption between agent hops — with real examples from production systems.

• Runtime verification without the rewrite. How to instrument your execution graph with lightweight trust checkpoints — behavioral tripwires that fire during execution, not in your test suite.

• Halt, retry, or escalate? A decision framework for real-time agent health that doesn’t tank your latency or annoy your users.

You’ll leave with a concrete verification architecture, a failure taxonomy you can use in your next incident retro, and serious second thoughts about that agentic pipeline you’re about to ship.