An agent run is more than a request.

The root run

Every investigation starts from a root agent.run span. It carries the stable run ID, service identity, environment, start and end time, outcome, and trace correlation.

• One run can contain multiple model calls
• A run can hand off to another agent or service
• The final outcome should distinguish success, failure, timeout, cancellation, and partial completion

Child activity

Child spans preserve the causal path through the run. Tracey uses them to explain critical-path latency, locate failures, compare prompt or model versions, and connect agent behavior to infrastructure.

• Model generation and embedding calls
• Tool, MCP, database, retrieval, and external API calls
• Routing decisions, retries, fallbacks, and handoffs
• Side effects and result classification

Evidence completeness

Tracey reports when telemetry is partial. A run with no tool spans cannot support a confident tool-failure explanation, and a run with no outcome cannot be treated as healthy.

Observed facts, deterministic calculations, and hypotheses are separate categories in Tracey’s evidence model.

Recommended attributes

Use stable, privacy-safe attributes for correlation. Keep prompts, responses, tool payloads, credentials, and personal data out of exported attributes by default.

tracey.run.id
tracey.agent.id
tracey.environment
tracey.outcome
tracey.prompt.version
gen_ai.system
gen_ai.request.model
gen_ai.usage.input_tokens
gen_ai.usage.output_tokens