The problem
Building a demo agent got easy. Running one in production did not. Teams that shipped agents hit two questions:
- How do we deploy it safely? A new prompt or tool can break an agent in ways a unit test misses.
- What is the agent allowed to do? It calls real tools with real data.
Scripts and manual checks work for one agent. They fail at ten.
The decisions
- One control plane to build, deploy and govern agents, so one team sees the whole agent.
- Governance by default. A platform team sets the rules once. Agent teams ship inside them.
What I left out
Another agent framework. Teams keep the framework they already chose.