agentic-production-readiness

Prepare an AI agent system for production operation. Cover SHIELD controls, sandbox/canary/production rollout, OpenTelemetry observability with GenAI semantic conventions, Agent Card drafting, governa

SKILL.md

GitHub

Agentic Production Readiness

Use this skill when an agentic system is close to deployment or operational handoff. Verify current protocol details instead of hardcoding unstable specs.

Workflow

  1. Identify deployment target, owner, data sensitivity, tool permissions, and rollback path.
  2. Apply SHIELD:
    • Separation of duties.
    • Human approval for high-impact actions.
    • Input/output validation.
    • Enforced security models.
    • Least agency and least privilege.
    • Defensive controls such as sandboxing and SCA.
  3. Plan staged rollout: sandbox, canary, then production. Define rollback triggers before release.
  4. Configure observability: logs, traces, metrics, token/cost tracking, model versions, tool calls, latency, and policy versions.
  5. Draft an Agent Card if external discovery or agent-to-agent use is required. Verify the current A2A location and schema before publishing.
  6. Generate DEPLOY_CHECKLIST.md from assets/templates/DEPLOY_CHECKLIST.md.

Governance Checks

  • Named accountable owner.
  • Credential source and rotation policy.
  • Memory retention and review policy.
  • Eval thresholds and release gates.
  • Incident response and audit trail.
  • Regulatory mapping only when the domain requires it.

References

Read ../agentic-engineering-sdlc/references/day5_prototype_to_production.md for Agent Ops, durable execution, and deployment patterns.

Done Criteria

  • Rollout and rollback are explicit.
  • Observability covers behavior, tools, cost, and safety.
  • Agent Card details are marked draft until checked against the current spec.
  • Governance and on-call ownership are assigned.