Let Me Be Your OpenTelemetry Champ
summary
A prototype OpenTelemetry MCP Server lets an AI agent inspect live telemetry, propose collector configs and upgrade paths, and automate tasks like automatic PII filtering and hands-free collector configuration. Demo uses coding agents to generate instrumentation PRs (initially using an outdated Python SDK 1.29 then corrected to 1.4.0), produce collector CRs, and detect missing receivers, but agents hallucinate components, emit wrong RBAC and OTTL, and require human review; overall the system works end-to-end but needs curated skills and an MCP-backed source of truth to be reliable.
AI-written summary · about
chapters (16)
- 0:00OpenTelemetry with AI agents
- 1:24Sprint to instrument demo app
- 2:50Manual instrumentation attempt
- 5:38Agents with specialized skills
- 7:23SDK and conventions in action
- 9:28Second round results
- 10:52Limitations of coding agents
- 11:33Deploying OpenTelemetry Collector
- 14:21Collector config and RBAC issues
- 15:27Upgrading collector breaking changes
- 16:32Removing PII from telemetry
- 17:36Verifying PII removal
- 18:57Summary of learnings
- 20:00Improving collector config
- 21:44Conclusion and future
- 22:48Q&A on agents and exporter
related talks
23mObserving the Observers: Bringing OpenTelemetry to Autonomous AI Agents
17mThe State of MCP observability: Observable.tools
27mKeynote: Tracing the Agent's Mind: Extending OpenTelemetry for Deep...
27mAccelerating Incident Response With OpenTelemetry and MCP: Building...
38mPractical introduction to OpenTelemetry tracing
33mWhat Did My Agent Do? Observability and Accountability for AI Agents
16mUsing the OpenTelemetry Collector to Empower End Users
30mOpenTelemetry 101