How the Agent Control Plane actually works
Most AI governance platforms describe policy. This is the mechanism underneath it — authority, runtime control, intervention, evidence, and decision provenance — each stage backed by a real endpoint you can call yourself, not a diagram of something that only exists in a sales deck. Every "see it live" link below goes to the actual, working thing.
Five stages, one action
What happens, in order, every time an agent asks the Control Plane whether it may act.
A key, not a claim
Every agent authenticates with its own registered key. Whatever identity the request claims is ignored — the key alone decides who's really acting, and it can't be spoofed from the request body.
Try it with a real key →Checked before it happens
The agent calls the Control Plane before the action executes and waits for allow / review / deny. It's a live policy engine consulted on every call, not a log line written after the fact.
Run a live evaluation →Held, not guessed
A review decision pauses the action until a person resolves it — approved, denied, or edited. A workspace-wide kill switch can halt everything instantly if you need a blunter stop.
Sealed, tamper-evident
Every decision — allow, review, or deny — is appended to a hash-chained ledger. Raw payloads are never stored, only a redacted copy. Edit one sealed record afterward and the chain breaks.
Verify the chain →Reconstructable, later
Every entry records which controls matched and why, and links to the entry before it. Months later, you can reconstruct exactly what happened and prove the record hasn't changed since.
See the recorded reasons →Live from the public demo ledger
This is real data from the same free, anonymous Workbench tool anyone can use — no account, no sales call. Run an action above and these numbers move.
—
What actually happens on the wire
Agent calls POST /control-plane/evaluate
With Authorization: Bearer <its key> and what it wants to do — a tool, an action type, a target, an optional payload.
Server resolves identity from the key
Looks up which registered asset that key's hash belongs to. If the key is unrecognized, suspended, or revoked, the call is denied right here — before any policy runs.
Policy engine decides
Checks the workspace kill switch and every active compiled policy. Sensitive payload values are redacted before anything is stored.
Decision returned and sealed
The agent gets allow, review, or deny, plus the matched controls and reasons. The same response is appended to the hash chain, linked to the entry before it.
If held, a human resolves it
A review decision waits. Approving or denying it writes a new ledger entry referencing the original — the original held decision is never edited or deleted.
Anyone can verify the chain, any time
Re-walking every hash from the beginning proves nothing in it has been altered — the same check this page's live numbers above are drawn from.
Go deeper
This page is the picture. The API reference, request/response shapes, and SDK snippets are one click away.