Proposal

Add opt-in OpenTelemetry instrumentation conforming to the OTel gen_ai.* semantic conventions, so downstream observability tools (Datadog, Honeycomb, Phoenix, Langfuse, OTLP collectors, Vercel's built-in observability) get structured visibility into agent execution. Off by default, zero runtime cost when disabled.

Context

The agent workflow currently emits useful logs and platform metrics but no structured spans. As background coding agents move toward production, teams need a standard wire format for tracing. The gen_ai.* conventions are the emerging standard, and AI agent workflows are an active focus area in the OpenTelemetry GenAI SIG (open-telemetry/semantic-conventions#2912).

Implementation

I prepared a branch with the change. It's six files / +174 lines:

apps/web/instrumentation.ts — Next.js server-side OTel registration via @vercel/otel, dynamic-imported only when OTEL_ENABLED=true
packages/agent/telemetry.ts — getAgentTelemetry() helper returning the AI SDK experimental_telemetry config
packages/agent/open-agent.ts — prepareCall now sets experimental_telemetry so all openAgent.stream() callers get instrumented automatically
packages/agent/index.ts — re-exports the helper
apps/web/package.json + bun.lock — adds @vercel/otel ^1.13.0
apps/web/.env.example — adds OTEL_ENABLED, OTEL_SERVICE_NAME, OTEL_EXPORTER_OTLP_ENDPOINT
README.md — new "Telemetry" section

Branch (passes bun run ci locally): https://github.com/meshailabs/open-agents/tree/feat/otel-gen-ai-instrumentation

Diff against main: https://github.com/vercel-labs/open-agents/compare/main...meshailabs:open-agents:feat/otel-gen-ai-instrumentation

Why I'm opening an issue

I tried to submit this as a PR but the repo is gated to collaborators only. Happy to clean up, scope down, or split the patch in any direction the maintainers prefer — let me know if there's a path you'd like to take here.

Activity

No activity yet.