Explore the product
A model call can succeed.
The task can still fail.
Understand the journey from a model response to recorded evidence before connecting your own agent.
Four moments that matter
Response → Investigation → Comparison → Verification
The experience is intentionally progressive. Open a stage to understand what to inspect—and what Tracelane does not claim.
01A successful API response is not a successful task
01 / RECOGNIZE
A model call can return HTTP 200 while the user's actual task remains unfinished. The response confirms transport, not the business outcome.
What you can inspect
- Read the recorded agent and tool steps, and the request text if your workspace has turned on input capture (off by default).
- Use an explicitly recorded success or failure outcome, rather than guessing from HTTP status.
Illustrative scenario. No customer trace or HTTP response is represented here.
02Follow the captured steps
02 / INVESTIGATE
Use Tracelane's trace tree, waterfall, per-agent lanes or transcript, then narrow to failures. Captured spans, rather than a generated story, are the evidence.
What you can inspect
- Inspect the relevant model, tool and agent spans.
- Check what was captured and what remains unknown.
The available information depends on actual capture settings and metadata.
03Find what changed without inventing a cause
03 / COMPARE
When an earlier run by the same agent has no recorded failure signal, the incident packet can show captured differences in model, tools, release and environment. Missing fields show as not captured.
What you can inspect
- Contrast the failure with the most recent comparable run that has no recorded failure signal. That does not prove the earlier run succeeded.
- On Builder and above, with input capture on and a role that can manage datasets, export the captured input as a regression fixture, labelled recorded or mocked. The file contains no assertions; you add your own.
A comparison is not always available; regression export does not rerun a live agent or its tools.
04Verify a bounded audit range
04 / VERIFY
Tracelane's offline verifier checks specific recorded audit evidence and signatures. It does not turn an incomplete capture into proof that every possible event occurred.
What you can inspect
- Export a date range, then check its hash chain and batch signatures against your trusted workspace key.
- Distinguish gateway-admitted ledger events from SDK/OTLP-only spans.
This walkthrough does not perform cryptographic verification or assert compliance.
Try it with your evidence
A real first trace is the next step.
Connect an agent using the documented quickstart, then inspect the actual captured evidence in your workspace.