How OpenTelemetry observes AI agents without code changes

Описание и применение

OpenTelemetry eBPF Instrumentation (OBI) позволяет наблюдать за трафиком ИИ‑агентов без изменений кода, обеспечивая согласованную телеметрию и диагностику проблем с помощью модели причинно‑следственных связей.

Задача

AI agents generate dynamic request paths that traditional request metrics miss, making it hard to attribute token usage and tool calls to specific agents, especially when SDK instrumentation is absent or fragmented.

Результат

OBI enables observation of AI agent traffic without code changes, providing consistent telemetry across languages and providers. When combined with a causal model, it allows teams to diagnose performance issues and attribute token usage to specific agents.

Рабочий процесс

• Deploy OBI as a DaemonSet on the cluster running the agents.
• Configure OBI to capture OpenAI, Anthropic, Gemini, Bedrock, Qwen, MCP over JSON‑RPC, embeddings, and rerank providers.
• Collect consistent telemetry using GenAI semantic conventions.
• Integrate OBI data with existing metrics, logs, and traces.
• Apply Causely’s causal model to match observed symptoms with known cause signatures.
• Review diagnoses to identify and address agent or infrastructure failures.

Доказательства

OBI attaches kprobes and uprobes in the kernel and parses protocols in a separate process outside your application.

Источник

One customer's token usage on AWS Bedrock and OpenAI grew 320% in a month, and the team could not tell which agent drove it.

Источник

As of OBI v0.9.0, coverage includes OpenAI, Anthropic, Gemini, Bedrock, Qwen, MCP over JSON‑RPC, and embedding and rerank providers.

Источник