Agent tracing
Runs that finish can still be wrong
Most agent failures do not raise an error. A tool retries quietly, a handoff stalls or the model takes a longer path, and the run still reports success.
Every span, with its cost and outcome
Each run is recorded as a trace you can read as a timeline or a tree. Spans carry latency and token cost, and the trace links to eval scores and to what the user did next.
Capture the run, read it in order, find the span that changed, and check the eval and user data around it.
Capture
Record agent turns, prompts, completions, tool calls, handoffs, retries, latency, and token usage.
Reconstruct
View the run in order so engineers can see which span or dependency changed behavior.
Diagnose
Connect the trace to eval failures, cost spikes, infrastructure events, and affected users.
Route
Send the incident to the owning team with the span, evidence, and downstream impact attached.
What teams can do
- Trace prompts, completions, model calls, tool calls, retries, and handoffs
- Spot slow spans, stalled handoffs, failed tools, and runaway loops
- Connect traces to eval scores, token cost, user impact, and business KPIs
- Filter and inspect high-cardinality production traces without losing context
Can Anosys trace multi-agent workflows?
Yes. Anosys captures agent turns, handoffs, model calls, and tool calls so teams can understand the entire run, not just the final response.
Is this only for OpenAI?
No. Anosys supports OpenAI, Claude Code, Anthropic, custom LLMs, OpenTelemetry, REST APIs, and SDK-based instrumentation.