Production operations

Production Operations

A production operating model for AI systems covering telemetry, SLOs, incidents, behavior, cost, change, recovery, and retirement.

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.