Use these artifacts after design begins—not only after launch. They define how a production-agent workflow is promoted, observed, contained, recovered, changed, reviewed, and retired.
| Need | Start with | Output |
|---|---|---|
| Decide whether to promote | Release gates | Evidence-backed hold, shadow, canary, bounded-production, or autonomy decision |
| Define runtime evidence | Telemetry contract | Trace, effect, version, identity, outcome, and cost events |
| Set service targets | SLO scorecard | Segment-specific objectives, budgets, capacity, and recovery policy |
| Contain and recover | Incident runbook | Scoped containment, reconciliation, regression, ownership, and re-enable decision |
| Change behavior safely | Change management | Compatible release bundle, evaluation report, canary, rollback, and post-change decision |
| Keep dependency views useful | Map freshness and change impact | Derived maps with source provenance, freshness, impact review, and no shadow authority |
| Detect dangerous divergence | Behavior monitoring | Independent intent/action signal routed to trusted containment controls |
| Govern capability provenance | Capability supply chain | Admitted, pinned, constrained, monitored, and revocable tools, MCP servers, skills, CLIs, and code packages |
| Measure adoption and transfer ownership | Delivery and adoption plan and customer handoff | Predeclared adoption contract, exercised harness ownership, and artifact lineage |
| Review production value and service health | Production service review | Expand, continue, constrain, pause, improve, or retire decision |
| Review an FDE or applied-AI portfolio | FDE and applied-AI portfolio review | Cohort-aware investment, continuation, productization, transfer, capacity, or exit decision |
| Route field learning | Field-learning register | Confidentiality-reviewed customer configuration, product backlog, reusable artifact, or retirement input |
| Improve or retire | Operate and Scale | Gated compatible release or verified decommission |
Start adoption instrumentation, artifact-lineage capture, and customer harness pairing during the pilot. The recurring cadence and improve/retire sequence are in Operate and Scale.