Loop Engineering Is a Control Plane Problem

Loop engineering is real. The idea: design the system that orchestrates your agents, not prompts you write once and forget.

Addy Osmani laid out the engineering primitives beautifully. But the article ends where the architecture decision begins. The pattern scales beyond coding: any autonomous loop acting on production systems faces the same architectural questions.

For those of us running agents on production enterprise systems, a loop touching regulated data, business-critical pipelines, and real users — without a human at the decision point — is not a tooling problem. It's a control plane problem.

Does the agent see the right data — scoped, current, and nothing it shouldn't? Who is accountable when the loop decides? Who prevents it from acting outside its authority?

The good news: we know what the challenges are. Identity. Lineage. Observability. Authorization. Harder than before, but named problems get solved.

The winners won't have the fastest loops. They'll have the best control plane. This is the layer I'm building for production enterprise systems, and the questions are harder than the tooling suggests — starting with which actions an agent should not be allowed to take on its own in the first place.