# Orchestration Trace

> A trace that records coordination events in a multi-agent system, such as spawning sub-agents, delegating work, communicating, aggregating results, and stopping.

- Category: Evaluation & Safety
- Stage: Emerging
- Also called: Multi-agent trace, Temporal interaction graph
- Canonical: https://www.joelcsaji.com/ai-dictionary/orchestration-trace

## Plain English

It is the map of how the agent team actually worked, not only a list of model calls.

## Example

A research supervisor's trace shows which specialist agents were spawned, which sources each used, and why the supervisor stopped the run.

## Why it matters

Multi-agent systems fail in routing and coordination, so teams need traces that expose those decisions for debugging, evals, and optimization.

## Related

- [Trace](https://www.joelcsaji.com/ai-dictionary/trace)
- [Trace Grading](https://www.joelcsaji.com/ai-dictionary/trace-grading)
- [Multi-Agent System](https://www.joelcsaji.com/ai-dictionary/multi-agent-system)

## Sources

- [arXiv: Orchestration Traces](https://arxiv.org/abs/2605.02801)
- [OpenAI: Trace Grading](https://platform.openai.com/docs/guides/trace-grading)