Add instrumentation for @openai/agents #16963

Activity

1. matejminar

   added a parent issue on Jul 11, 2025

   • Add Agent Monitoring Support #16960
2. nicohrubec

   marked Support agents in openai integration #17347 as a duplicate of this issue on Jan 5, 2026
3. stephanie-anderson

   added

   Feature

   on Mar 24, 2026
4. andreiborza

   removed a parent issue on Apr 29, 2026

   • Add Agent Monitoring Support #16960
5. andreiborza

   added a parent issue on Apr 29, 2026

   • Agent Monitoring Backlog [JavaScript] #20576
6. mydea commented on Sep 18, 2026

   

   mydea

   on Sep 18, 2026

   Member

   More actions

   Scope + relevance (adding detail to this request)

   Relevance (npm)

   • @openai/agents: ~6.3M downloads/month (~33.3M last year) — running at more than 2x the trailing-12-month average, i.e. steep growth. First-party OpenAI agent framework, and unlike most OpenAI-compatible providers it is not covered by any of our existing proxies (Vercel AI SDK / OpenAI client / LangChain) — it's its own orchestration runtime, so agent runs, handoffs, and tool calls are currently invisible.

   Rough scope

   • Add instrumentation logic under packages/server-utils/src/ai/openai-agents/ and wire it up via src/integrations/openai-agents.ts, registered in getTracingIntegrations(); runtime packages re-export from @sentry/server-utils.
   • The SDK (openai-agents-js) exposes a tracing/hooks system — prefer hooking that over monkey-patching. Map its concepts to gen_ai.* spans per the span attribute conventions: agent invocation (invoke_agent), model/LLM generations (chat), tool executions (execute_tool), and handoffs, with the run as the parent span.
   • The underlying model calls go through the OpenAI client we already instrument — make sure the agent spans nest correctly with (and don't double-count) those, similar to how we handle LangChain/LangGraph over the provider SDKs.