Observing the Observers: Bringing OpenTelemetry to Autonomous AI Agents
summary
Telemetry is blind to agent identity and authorization so attackers exploit MCP servers: 206,000 attack events recorded in honeypots with 75% targeted MCPs and 45% repeat attackers. The proposal instruments agents with OpenTelemetry spans that carry agent ID/public key, capability, drift and trust scores, plus a child decision path recording allow/deny and step-level outcomes. Working demos push traces into Tempo and show open-source adapters for LangChain, CrewAI and custom agents plus a reference implementation and OTel issue numbers to review.
AI-written summary · about
chapters (8)
- 0:00AI agent security challenges
- 2:26Telemetry blind spots
- 3:52Standards in AI agent ecosystems
- 7:01Authorization attributes proposal
- 10:15Detection with trust scores
- 13:26Authorization observability demo
- 18:43Standardize AI agent telemetry
- 20:33Q&A
related talks
25mLet Me Be Your OpenTelemetry Champ
31mDeep Dive To AI Agent Observability
18mWhat did AI just do? Making LLMs Observable with OpenTelemetry
33mWhat Did My Agent Do? Observability and Accountability for AI Agents
29mNon-Invasive AI Agent Observability With OBI
18mOpenTelemetry and Observability: What, Why, and Why Now?
17mThe State of MCP observability: Observable.tools
27mObservability as a foundation for AI-enabled platforms