Agent trace semantics for real-world AI workflows
An early-stage concept note scoping how agent trace semantics should extend to delegation, hand-off, and multi-agent coordination beyond the single-agent case.
Whether the span taxonomy developed for single-agent tool use (ZAN-RN-002) still attributes failures correctly once workflows involve delegated sub-agents, shared state, and asynchronous handoffs.
- Delegation span taxonomy
- Multi-agent failure attribution model
- Comparison against single-agent trace baseline
Instrument a toy multi-agent workflow with a planner and two delegated workers, capture delegation-scoped spans, and test whether a reviewer can attribute a failure to the correct agent without access to raw payloads.
Delegation-scoped spans, distinct from tool-call spans, are required to reconstruct failures in multi-agent workflows without collapsing separate agents' decisions into a single trace.
Why single-agent trace semantics are not sufficient
ZAN-RN-002 defined a span taxonomy for tool-using agents. Real-world deployments increasingly delegate subtasks to other agents, which introduces handoff boundaries, shared and forked state, and asynchronous completion that the single-agent model does not represent.
Related research
View allTrace semantics for tool-using agents
What trace semantics are needed to explain failures in tool-using AI agents?
A measurement model for AI deployment readiness
How should teams measure whether an AI system is ready for controlled production deployment?
Reliability beyond pass rate
Which metrics better capture agent reliability than a single task success rate?