AI Agent Orchestration Platform · View 03 of 32 · 1 · Context and scope
The decision
- The control plane owns definitions, configuration, policy, identity, versions and governance. It never runs agent code
- The execution plane runs agents and workflows. It holds no authoritative definition, only a pinned copy of one
- The interface between them is a signed, immutable bundle — not a database both sides write to
Why it is worth the extra hop
- Execution scales on queue depth; the control plane scales on people and CI, which are three orders of magnitude apart
- Workers can be deployed per region for residency without moving the registries
- A registry deployment cannot change the behaviour of a run that is already in flight
Consequences accepted
- A control-plane outage blocks new admissions; in-flight runs continue on their cached bundle
- Two deployment pipelines and two scaling models to operate rather than one
- A definition change reaches running workloads only at the next admission, never mid-run
