Enterprise AI Agent Control Plane: Architecture and Controls

An agent that answers a question is easy to place in an existing application. An agent that keeps working after the conversation ends is harder to contain. It needs memory of the assignment, access to tools, a place to run, a way to report progress and rules for decisions it may make without asking. At that point, agent capability is no longer just a model feature. It starts to look like an operating layer for work.

OpenAI’s 29 September 2026 introduction of dots makes that shift concrete. The company describes persistent agents with their own cloud computers, connected applications and the ability to work on several projects over time. Users can inspect activity and set action rules. OpenAI also previewed specialist dots with separate organizational identity and credentials for defined responsibilities. The initial consumer and business rollout is limited to eligible plans and markets; the specialist model is a preview, so enterprise architects should distinguish the announced direction from generally available production capability. Source: OpenAI announcement.

The product demonstration is compelling because it joins capabilities that enterprises often buy or build separately. A long-running task can encounter new evidence, update an artifact and return with a decision for review. The agent can use applications where work already happens. That changes the comparison set for every agent platform: a workflow designer and a tool catalog are useful, but they are no longer the whole proposition.

The real architectural question is how authority follows the agent. A personal agent may operate with a user’s connected applications. A specialist agent working for a procurement or support function needs a distinct identity, constrained credentials and ownership that survives staff changes. A bank will also ask whether an action can be traced to the initiating request, the specific policy version, the data consulted and the approval that released it. A simple transcript is rarely enough to reconstruct a consequential business decision.

OpenAI says dots’ background research uses read-only access to connected apps, while more consequential actions go through rules and review. Those are sensible boundaries. They do not remove the need to test integration-specific behaviours: a tool labelled “read” can still expose sensitive data, and a sequence of individually permitted actions can produce an unintended outcome. Banks will need controls at the level of the business operation as well as the individual API call. Source: OpenAI announcement.

For ArchyAgents, this announcement sharpens the product position. Competing on a generic promise that agents can execute multi-step tasks will be difficult. A more defensible offer is an enterprise control plane that can coordinate several models and systems under explicit policy, particularly in a customer-controlled Azure environment. That means making identity, workflow state, tool permissions, human decisions, audit evidence and operational telemetry visible as first-class objects. Jira can remain a valuable use case, while the orchestration layer applies to broader work.

That is a strategic inference, not a claim that OpenAI cannot satisfy enterprise requirements. Its specialist preview may mature quickly. The question for a prospective customer is where they need independent control: model choice, deployment boundary, data access, integration with existing bank identity and risk systems, and the ability to change vendors without rebuilding every workflow. A product should demonstrate those controls in a real process rather than describe them on an architecture slide.

One useful test is a small change request that spans a source repository, Jira, documentation and an approval step. Let the agent prepare a modification, run checks and explain the evidence. Then interrupt it, change the requirement, revoke one permission and inspect what happens. If the system can show why it stopped, what it retained and what it would do next, it is beginning to behave like an accountable operating layer.

Watch the availability and APIs for specialist identity, audit trails, permissions and cross-system orchestration. Those details will decide whether persistent agents are primarily a convenient personal experience or a foundation that enterprises can govern at scale.

For the related architecture decisions, see AI Agent Governance: Identity, Authority and Accountability; Enterprise AI Workflows: State, Approvals and Recovery; AI Agent Observability with OpenTelemetry.