Building the Production-Grade Platform for Enterprise AI

What happens when your first AI agent becomes five hundred, and how to build a platform that keeps them safe and fast to ship. Sergey Kurson opened with why agent platforms matter now (agent sprawl, shadow AI, integration chaos, cost blindness), the build, deploy and operate stack on open protocols (MCP, A2A, OpenTelemetry), a four-level platform maturity model and the federated operating model.

My part was the deep dive, framed as three questions for every agent in production: can you control it, can you see it, can you change it. The controls live outside the agent. A governance plane declares policy: Cedar with default-deny, an identity per agent, and a registry as the only path to credentials. An enforcement plane applies it on every request through four stop points: inbound gateway, guardrails, human-in-the-loop and outbound gateway. Then observability at the level of the whole agent trajectory, offline, on-demand and online evaluation, agent CI/CD where the artifact is code, prompts, model version, tool schemas, policy and state, and the platform side: hard budgets, per-agent cost attribution and tenant isolation down to the vector store.

Co-presented with Sergey Kurson, Principal Solutions Architect at AWS, on Stage 2 (Solution Track) of the first Tech Race Summit by SOFTSWISS at Centrum Praskie Koneser, Warsaw.