Agent architecture needs a durable source of authority.
When clients, agents, and models can all influence a workflow, the system needs one place where policy, authorization, state transitions, and acceptance remain authoritative.
The control plane owns decisions that must survive model variability.
Models are probabilistic and clients are replaceable presentation surfaces. Neither should silently become the system of record for permissions, tenant boundaries, durable job state, or release decisions.
A control-plane architecture centralizes identity context, policy, workflow state, capability contracts, approvals, validation, evidence, and recovery rules behind stable interfaces.
Agents become bounded capabilities.
Specialized agents can still be powerful. Their role is clearer when they operate through explicit inputs, outputs, permissions, allowed callers, allowed targets, and escalation paths instead of relying on human-readable names or prompt conventions for authority.
Clients remain projections.
Desktop, mobile, and web clients can evolve independently when they consume authoritative backend state rather than recreate policy locally.