What an AI agent control plane should manage: Axon's operating view
An AI agent control plane cannot be just another chat box. Without an operating view, teams waste hours every week chasing repetitive status updates, checking files manually, and finding errors only after a customer brief or internal report has already moved forward. OpenAI’s Codex app introduction discusses Skills, background work, and a foundation for broader knowledge work, which signals a shift from conversation interfaces toward work operations. For an AI workforce, the control plane must show queues, state, cost signals, approvals, and artifacts. See Introducing the Codex app.
Chat is not enough for operations
Chat is useful for temporary questions. It is weak for managing many AI digital employees at once. A team may run research summaries, contract cleanup, email drafting, spreadsheet updates, and customer follow-up in parallel. Without an AI agent control plane, owners search conversation history to learn whether a task is waiting for inputs, actively running, failing, consuming resources, or waiting for approval.
Axon’s operating view is centered on workflows rather than chat logs. Skills define available capabilities. Agents define execution plans. The workspace stores deliverables. Trust Mode protects risky actions. Teams that need the capability layer can start with the System Skills introduction. Teams ready to assemble a process can use AI Build for the first Agent.
The control plane is not about making AI look more autonomous. It is about helping people see who is doing what, where work is blocked, and whether the next action is safe.
Minimum run fields for the control plane
The foundation of an AI agent control plane is a searchable record for each run. The following model is enough to make an Axon operations view practical.
{
"runId": "customer-brief-2026-05-22-09",
"agent": "customer-brief-agent",
"status": "waiting_for_approval",
"inputOwner": "sales-ops",
"skillsUsed": ["web-research", "pdf-read", "email-draft"],
"artifacts": ["sources.md", "brief.docx", "email-draft.md"],
"riskLevel": "medium",
"approvalRequired": true,
"nextAction": "sales manager confirms recipient and wording"
}
- Step 1: create a unique runId so multiple conversations do not blur together.
- Step 2: record the input owner to diagnose missing fields and old source versions.
- Step 3: record Skills used so failures can be traced to capability or orchestration.
- Step 4: move artifacts into the workspace instead of leaving them in chat.
- Step 5: show risk level and next action to the workflow owner.
Three views: queue, state, artifacts
Queue view
The queue view answers which task should run first. It should show source, priority, due time, and input completeness. For recurring work, it should also distinguish scheduled triggers, manual triggers, and reruns.
State view
The state view answers where the task is blocked. Useful states include queued, running, waiting_for_input, waiting_for_approval, failed, completed, and archived. Scheduled jobs especially need transparent state; see the scheduled Agent manual verification guide.
Artifact view
The artifact view answers whether the output is reviewable. Each Agent run should place files, summaries, source lists, drafts, and reviewer notes in the workspace. If the task sends email, publishes content, or affects external systems, route it through the Trust Mode email confirmation guide.
| Control plane view | Operations question | Axon object |
|---|---|---|
| Queue | Which task runs first? | Agent run queue |
| State | Is it blocked on input, execution, or approval? | Run status |
| Artifacts | Can the output be accepted? | Workspace artifacts |
| Risk | Does a person need to confirm? | Trust Mode |
| Review | Why did it fail or rerun? | Run record |
The useful detail is not the table itself; it is the operating habit behind it. A manager should be able to open one run, see the brief, inspect the files, understand the blocking state, and decide whether the next action should continue. That keeps the system from becoming a collection of invisible automations. It also helps the team compare similar runs over time: which customer briefs always wait for missing inputs, which research jobs consume too much review time, and which Skills should be promoted into a standard workflow.
FAQ
Q1: How is an AI agent control plane different from a task manager?
A task manager records what people should do. An AI agent control plane also records capabilities used, files created, automation state, and risk approvals.
Q2: Does a control plane mean every step needs human supervision?
No. Low-risk steps can run automatically. High-risk actions require confirmation. The control plane makes the boundary visible.
Q3: Where should cost be tracked?
Cost should attach to runId, Agent, input size, and capability used. That lets the team decide whether a task deserves long-term automation.
Q4: Do small teams need a control plane?
Once a team runs multiple Agents or recurring jobs, it needs at least a minimal control plane. Otherwise status lives in chat logs, folders, and personal memory.
Next step
Get started in Axon by adding runId, status, artifacts, and approvalRequired to one existing Agent. Then learn more about Skills and use the AI agent control plane to manage digital employees, not just watch chat output.