Skip to main content

Crate agenttrace

Crate agenttrace 

Source
Expand description

Cost + latency aggregation for LLM agent runs.

Groups LLM calls into named runs (e.g. one user request, one agent turn) and emits per-run + global stats: total cost, p50/p95 latency, per-model breakdown. Composes with cachebench — feed it CacheTracker metrics and you get the full observability stack.

§Quick start

use agenttrace::Tracer;
use std::time::Duration;

let tracer = Tracer::new();
let run = tracer.run("answer_user_question");
run.record("claude-sonnet-4", 0.0042, Duration::from_millis(1230));
run.record("claude-sonnet-4", 0.0017, Duration::from_millis(540));
run.finish();

let agg = tracer.aggregate();
assert_eq!(agg.runs, 1);
assert!(agg.total_cost_usd > 0.0);

Structs§

Aggregate
Cross-run aggregate stats.
CallRecord
One LLM call captured under a run.
ModelBreakdown
Per-model rollup.
RunHandle
In-flight run handle. Drop or call finish to seal it.
RunRecord
One completed run: a name plus the calls it contains.
Tracer
Records runs and exposes aggregate stats.