How we monitor internal coding agents for misalignment

OpenAI describes the monitoring infrastructure they built to detect misalignment in internal coding agents that have access to real systems, documentation, and even their own safeguards. The post covers behavioral telemetry, risk factors unique to self-referential agent deployments (agents that can inspect or modify their own guardrails), and lessons learned from real-world agentic workflows. Directly relevant to anyone shipping or operating autonomous coding agents at scale.