Profile validation
Schema, required fields and format. Catches structural errors before anything runs.
Agent Studio
Launch and Promote
A readiness checklist stands between a draft and your customers. Every check must pass, a named person must sign off and the agent starts in shadow mode. Autonomy is earned, not granted.
Launch sequence
Four steps from draft to live.
Every launch follows the same path and leaves the same evidence. Your compliance team reviews one process, not a different story per agent.
Pass readiness
All checks must pass: valid profile, resolved tools, attached guardrails, evals above the bar and a passed contract negotiation.
Owner sign off
The owner reviews the readiness report and signs off. The decision is recorded on the ledger with the signer and the reason.
Deploy at shadow
The agent starts in shadow mode. It sees real traffic but a person acts. The agent earns its way up from here.
Promote with evidence
Promotion from shadow to assisted to autonomous requires eval evidence and a two-person request. The platform records the promotion as a signed git tag.
Promotion path
Shadow, assisted, autonomous.
Every agent starts restricted. It earns more freedom through eval scores and promotion requests. The platform enforces the ceiling.
Shadow
The agent runs alongside a human. It sees real inputs and produces outputs, but a person acts. Every step is recorded for comparison.
Assisted
The agent handles read operations on its own. Writes and escalations go to the approval queue. A named person decides every action that changes a record.
Autonomous
The agent handles reads and approved categories of writes on its own. Exceptions and novel situations still escalate. Earned through evals and a two-person promotion request.
Readiness checks
One list that says go or not.
The readiness report is a single view of everything the agent needs before launch. It updates in real time as you fix issues.
Tool resolution
Every tool resolved against the connector hub. Missing connectors block the launch.
Guardrail coverage
Domain guardrails attached and universal guardrails present. Coverage gaps are flagged.
Eval gate
Eval scores above the configured threshold. Below the bar means the agent cannot launch at full autonomy.
Contract negotiation
The profile meets the runtime contract. Tool classifications, guardrail types and autonomy levels all checked.
Owner sign off
A named person reviews and approves. The decision is on the ledger with signer, reason and timestamp.
Version control
Every change is a new version.
Git-backed profiles
The AOP lives in git. Every save is a commit. Promotion is a signed tag. You can diff, review and roll back.
Version history
See what changed, who changed it and when. Compare any two versions side by side.
Audit trail
The readiness report, the sign off decision and the promotion request are all recorded on the hash-chained ledger. When an auditor asks who approved a launch and what evidence they reviewed, you point at the ledger.
FAQ
Common questions about launch and promotion.
No. Every agent starts at shadow. The platform enforces this. Promotion to assisted requires eval evidence. Promotion to autonomous requires a two-person request and higher eval scores.
The platform clamps the agent to assisted mode at runtime. It cannot run at full autonomy until it passes the eval gate again. This is enforced on both the contract and engine paths.
Every save creates a new version. The AOP is a git-backed YAML file. You can diff any two versions, see who changed what and roll back to any previous version. Promotion is a signed git tag.
Yes. The version history shows every change. You can roll back to any previous version and the platform will enforce the older profile immediately. The rollback is recorded on the ledger.
See the launch flow end to end.
We will walk through readiness checks, sign off and the first shadow run together.