Keep the SDK in charge. See the whole run.

Sulcus instruments the SDK’s Runner calls and lifecycle hooks. The SDK still owns the agent loop, tools, handoffs, sessions and guardrails.

Hosted by Sulcus, SDK 0.22.2+

Sulcus adds

A runtime view across agents.

  • Agent lifecycle, including handoffs between agents.
  • Model calls and local tool activity.
  • Guardrail events when a guardrail triggers.
  • Reported token usage and an optional per-run token limit.
  • Stop control and a bounded event history for each run.

The SDK keeps

How agents behave.

  • The agent loop, model selection, tools and handoffs.
  • Sessions, approval state, SDK limits and results.
  • Hosted tools and provider behavior outside local hooks.
  • Anything your application stores, and its side effects.

How it attaches

Around your normal Runner calls.

When a project selects the OpenAI Agents SDK, Sulcus activates its instrumentation in the run’s container and wraps Runner.run and Runner.run_sync. Your application calls the SDK as usual. Sulcus supports SDK 0.22.2 and later 0.22 releases; Runner.run_streamed isn’t instrumented.

Observation boundary
Runs the agents
OpenAI Agents SDK
Instrumented calls
Runner.run / run_sync
Event source
Lifecycle hooks and tracing spans
Not collected by instrumentation
Prompts, model outputs, tool payloads

What this integration doesn’t do.

If your agents request tool approval, the request can appear in Run Details. For hosted runs, Sulcus can’t approve, deny or resume it. What your app prints is kept with the run, so avoid printing sensitive data.

Provider-hosted tools, internal retries and nested runs outside the instrumented Runner call don’t get full lifecycle coverage. The token limit applies per run, using reported usage.

See your first run in Sulcus.