Engineering Production Agentic Systems, Part 2: Guardrails

MCP gateways have made wrapping a hundred internal APIs as agent-callable tools a one-day, sometimes one-hour, job. But exposing tools cheaply isn't the hard part — ensuring correctness is: making sure an agent picks the right tool, with the right parameters, under the right authorization, with the right human review when needed.

This is Part 2 of a three-part field manual on engineering production agentic systems. Part 1 covered context engineering; this part covers guardrails — specifically tool surface design, authorization scopes, and audit-trail engineering. Part 3 will address human-in-the-loop topology.

The underlying thesis: production agentic systems succeed or fail based on these architectural disciplines, not model choice. Tool correctness is the unglamorous work that never shows up in agentic-AI demos, yet it's precisely where production systems either hold up under real-world load or quietly collapse.