AI Agent governance

Deploying AI agents in production exposes gaps in traditional safeguards designed for single LLM requests. Agents chain model calls, invoke tools, and trigger external side effects across multiple steps, making per-request controls insufficient. Five governance pillars are outlined: execution control (human oversight insertion points), tool and action permissions (runtime-evaluated, discovery-level access control via a central MCP registry), cost and resource governance (session-level budget enforcement), and policy enforcement at every action point rather than just the LLM boundary. The recommended approach is enforcing these controls at a shared infrastructure layer — an AI gateway sitting between agents and the systems they call — so governance applies consistently without modifying agent code. Portkey's AI Gateway is presented as an implementation of this pattern.