{
  "id": 15665654,
  "title": "How OpenTelemetry observes AI agents without code changes",
  "url": "https://urgent.news/2026/10/19/how-opentelemetry-observes-ai-agents-without-code-changes",
  "topic": "ai",
  "section": "AI",
  "published": "2026-10-19T22:42:16.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/causely/how-opentelemetry-observes-ai-agents-without-code-changes-bm3"
  },
  "original_language": "en",
  "account": "AI agents pose challenges for operations due to the rapid and unpredictable nature of their requests. Unlike traditional requests that follow a static path, AI agent requests dynamically expand into various components like embeddings, vector search, reranking, tool calls, and other agents. This dynamic nature makes it difficult to identify which agent is responsible for increased resource usage, as exemplified by a customer's token usage on AWS Bedrock and OpenAI soaring by 320% in a month. Traditional instrumentation struggles to keep up with AI agents, as coverage is fragmented across different programming languages and providers. OpenTelemetry's eBPF Instrumentation (OBI) addresses this gap by attaching kprobes and uprobes in the kernel and parsing protocols in a separate process outside the application, allowing it to observe LLM and MCP calls over TLS without code changes. OBI is deployed as a DaemonSet, providing one process per node without the need for application restarts. It offers comprehensive coverage of popular AI providers, including OpenAI, Anthropic, Gemini, Bedrock, Qwen, MCP over JSON-RPC, embedding, and rerank providers. OBI complements existing SDKs by capturing LLM and tool calls from outside the application process, as well as TLS, DNS, and retry information that SDKs cannot observe. Although consistent telemetry is achieved through the same attributes and semantic convention version for GenAI calls and trace context for connecting calls across services and languages, a gap remains in causal context. Causely fills this gap by building upon OBI and incorporating existing metrics, logs, traces, and infrastructure data to analyze observed symptoms against known cause signatures, including AI model-specific causes. This causal model provides diagnoses for GenAI applications, which are notoriously difficult to operate. To start using OBI, watch the full talk and deploy it as a DaemonSet on a cluster running your AI agents. The project is actively maintained and welcomes contributions through SIG calls or by addressing open issues in the OBI repository.",
  "summary": "Originally posted to causely.ai by Endre Sara AI agents are hard to operate because instrumentation can't keep up with them. Haibin Zhang of Alibaba Cloud and I made that case at KubeCon China 2026, and showed how OpenTelemetry eBPF Instrumentation (OBI) closes the gap. The full talk is at the end of this post. Multi-agent applications fail in ways request metrics miss A traditional request…",
  "key_points": [
    "OpenTelemetry's eBPF Instrumentation (OBI) observes AI agents without code changes.",
    "OBI deployed as DaemonSet provides one process per node, no application restarts.",
    "OBI complements SDKs by capturing LLM, tool calls, TLS, DNS, and retry info."
  ],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}