As AI agents move from generating code to testing, deploying, and operating it, the important question is no longer “How autonomous is the agent?” It’s “What can this agent safely decide without a human?”

Traditional delivery pipelines establish confidence through predefined tests and gates. But these only catch failures someone anticipated. Agentic systems introduce a harder problem: agents can change code, tests, and deployment processes themselves, creating failure modes nobody thought to test for.

This changes the role of observability. Monitoring answers questions we knew to ask; observability provides the evidence needed to investigate questions we didn’t know we would have. OpenTelemetry can provide the instrumentation, but instrumentation alone doesn’t guarantee this capability.

Borrowing from industrial automation, this talk introduces an as-designed vs. as-built model for agentic engineering. The specification describes what should exist; production telemetry provides evidence of what actually exists. Observability continuously reconciles the two.

Through a worked example, we’ll see an agent encounter an unexpected failure, investigate the running system, remediate or update the specification, and escalate when the evidence is insufficient.

The goal isn’t a “dark factory” operating without people. It’s an ambient factory: a platform that acts autonomously when the evidence supports it, and knows when to ask for help.