The Agent Failure Playbook: Production Gone Wrong

Production AI agent failures don't look like software bugs. They're subtle, expensive, and often invisible until the damage is done.

An agent loops a tool 47 times and burns $200 in a single request. An agent confidently calls an API with hallucinated parameters and corrupts downstream data. A context window silently overflows and the agent starts ignoring its own instructions mid-conversation.

These aren't hypothetical. They're patterns from building production AI at Microsoft, teaching agents to thousands of developers, and collecting community war stories from 600K+ followers asking "why did my agent do this?"

This talk is the failure playbook nobody writes until something goes badly wrong.

The failure taxonomy 5 production failure modes across three categories:

Reasoning failures - An agent overrides its own system prompt mid-conversation because a tool response contained conflicting instructions. No exception raised. No error log. Just wrong behavior.

Infrastructure failures - A retry loop with no exit condition hits an external API 47 times in 4 seconds. The agent reports success. The bill arrives later.

Integration failures - An agent constructs a valid-looking but hallucinated API payload. Pydantic doesn't catch it because the types are correct. The downstream database does.

For each failure: what it looks like in production, a real example, Python patterns that prevent it, and detection strategies.

Behind every failure is an architectural decision that seemed reasonable at the time. This talk gives you the production hindsight to make better decisions upfront