Multi-Agent Observability

For agentic systems, capture every sub-agent, tool, and retrieval step as typed child spans under one root trace.

For agentic systems — Google ADK, custom orchestrators, or anything built on a supported provider SDK — init() captures every sub-agent call, tool invocation, and retrieval step as typed child spans under one root trace, with no manual span-per-agent wiring.

Google ADK / OpenAI Agents SDK / CrewAI — built on the provider SDKs, so calls are captured through the LLM Providers instrumentation.
Because spans are W3C-compatible, a multi-service agent (e.g. an orchestrator calling out to separately-deployed sub-agents over HTTP) still stitches into one trace — see Distributed Tracing.

To stress-test a multi-agent system’s behavior end-to-end rather than reading traces after the fact, see Scenario Simulation — it drives your real agent through synthetic multi-turn conversations and scores the result.