What is Span?
Agent InternalsOne step in a traced run, recorded with its kind, inputs, outputs, latency, and cost. Span kinds in agentic tracing include model calls, tool calls, retrievers, rerankers, guardrails, evaluators, and handoffs.
Why It Matters
A span is the atom of agent observability. Every claim you can make about a run (where it went wrong, what it cost, which decision it made) is only as true as the spans underneath it. No span, no evidence; the step happened or did not, and nobody can say which.
The vocabulary comes from OpenTelemetry’s GenAI conventions, which means agent spans land in whatever observability stack a team already runs instead of a bespoke dashboard.
The Span Kinds
Model call. One LLM invocation: the prompt in, the completion out, tokens and cost.
Tool call. A tool-call record: the function requested, the arguments, the result, latency.
Retriever. A fetch from the index: the query and the candidates it returned.
Reranker. The re-ordering pass over retrieved candidates before they enter context.
Guardrail. A policy check that ran: what it evaluated and what it decided: allow, deny, or escalate.
Evaluator. An eval scoring attached to the run: rubric, score, judge.
Handoff. A task transfer between agents: what crossed the boundary and to whom.
Where It Breaks
Spans fail by absence. A step that was never instrumented is a hole in the trajectory, and incidents have a way of happening in exactly the uninstrumented parts. The second failure is spans without attribution: latency and cost recorded nowhere means the run’s expense is discovered on an invoice, not in a trace.
Sampling is the quieter break. Dropping telemetry to save cost deletes precisely the unusual runs that would have needed inspection.
How Flytebit Handles It
Full-trajectory tracing is the first thing we instrument in an observability engagement: every span kind captured, per-call cost attributed, nothing sampled away. The reasoning behind trajectory-first observability is in Observability for Agentic AI Governance.