Governed agent lifecycle

Agent Studio

Design agents as versioned operational products. Every prompt, model, tool, guardrail, evaluation and release decision remains inspectable.

Registry foundation ready

Lifecycle4 gatesDraft to approved release

Runtime defaultBlockedEnable only after a release

Write accessApprovalUnsafe definitions are rejected

Emergency control< 1 actionOperator kill switch

Workspace registry

Purpose-built agents

Initial ServiceNow pack roles shown as the intended product configuration.

Observe

Incident triage

Builds an evidence timeline and proposes classification, impact and next-best action.

ServiceNow readKnowledge readTelemetry read

Recommend

Resolution composer

Drafts work notes, stakeholder updates and resolution evidence without external writes.

Incident contextEvidence graphDraft output

Approval required

Runbook operator

Executes only allowlisted low-risk actions after policy and human approval checkpoints.

Approved MCP onlyPreconditionsPost-validation

Promotion path

Release lifecycle

Fail closed

Draft

Definition and tool permissions are content-addressed.

Review

A second administrator validates risk, budget and policy.

Evaluate

The declared golden suite must clear its release threshold.

Release

An approved version is promoted with idempotent evidence.

Release contract

Evidence gates

Definition integritySHA-256

Separation of dutiesRequired

Golden evaluationEnforced

Model cost ceilingPer run

Write toolsApproval only

Runtime safety

Global agent stop

Operators can stop execution immediately. Only owners or administrators can re-enable it.

Disabled until release

Release invariantNo agent executes without an approved version, passing evaluation evidence and an enabled runtime control.define → review → evaluate → release → enable → observe