> workflow / monitoring

Monitoring

Observe production behavior, model quality drift, and incident signals with actionable alerts — not vanity dashboards.

What should I do next?

Complete the steps and evidence below, clear the human gates, then continue to Production.

When to use

  • Service is in stage/prod and needs SLOs on the glass
  • Agent/tool paths lack tracing
  • You need alert routes and on-call runbooks before GA traffic
  • Eval drift or cost anomalies must page someone

Step-by-step

  1. Confirm golden signals and AI-specific metrics (latency, errors, tokens, tool failure rate).
  2. Instrument tracing across agent tool calls and model providers.
  3. Build dashboards that map to SLOs and abort criteria — delete unused charts.
  4. Define alert policies with runbooks; eliminate pages without owners.
  5. Wire Sentry/OTel (or equivalent) for errors and traces.
  6. Rehearse an incident with incident.prompt; fix gaps in paging.
  7. Schedule drift monitors for evals and prompt versions.

Decisions to make

  • What pages vs tickets vs dashboards-only
  • Retention and PII redaction in logs/traces
  • SEV definitions for AI quality regressions
  • Who owns model-provider outage response

Evidence to collect

  • Dashboards linked from runbooks
  • Alert policies with owners
  • Tracing enabled on agent/tool paths
  • On-call runbooks for top failure modes

Role-specific next actions

  • PM: Define customer-impact language for SEVs and comms triggers.
  • Architect: Ensure telemetry covers new boundaries and third-party deps.
  • Engineer: Add instrumentation with features; fix noisy logs.
  • DevOps: Own alert routing, retention, and on-call integrations.

Inputs

  • SLOs and abort criteria
  • Existing dashboards and log pipelines
  • On-call roster and escalation paths

Outputs

  • Production dashboards
  • Alert policies and runbooks
  • Drift monitors for evals/prompts

Required prompts

Required skills

Required MCP servers

Tooling & handbook

Recommended agents & patterns

Human approval gates

  • Incident commander required for SEV-1/2
  • Alert changes that mute pages require dual review

Automation opportunities

  • Tracing for agent tool calls by default
  • Eval drift monitors on prompt/version changes
  • Auto-open incidents from SLO burn alerts

Enterprise best practices

  • Retention and PII controls on prompts/logs
  • Mandatory post-incident reviews with actions
  • Vendor outage playbooks for model providers

Common mistakes

  • Alert fatigue from unowned pages
  • No tracing on agent/tool paths
  • Dashboards without SLO linkage