Follow an execution across its moving parts
A single agent task can cross a model, several tools, another agent, and a human decision. UpTrain is designed to bring those observations into one ordered record, with stable execution identity and the context needed to interpret each action.
The goal is a useful explanation of the run: what was requested, what the capture integration observed, how the work progressed, and where it stopped. A successful workflow status is kept separate from the quality of its output.
- Tool and model activity, services, and observed domains.
- Errors, retries, interruptions, recovery, and agent handoffs.
- Human approvals and declines at the point they happened.
- Observed payment states, without inferring settlement from a request.
Move from a fleet view to one specific step
The proposed activity view groups runs by agent, tool, outcome, and time. Reviewers can move from a pattern—such as repeated lookup failures—to the underlying execution.
Elapsed duration and captured activity time answer different questions. Concurrent steps may overlap, and missing measurements must remain unknown rather than being added as zeroes.
Keep the evidence class visible
Runtime capture and assistant-reported checkpoints have different sources. UpTrain’s intended model separates those classes in both individual records and aggregate metrics. A session summary from an assistant is useful context, but it is not proof that an external action occurred.
Preserve context beyond the interface
We plan structured exports, explicit capture coverage, and evidence-integrity workflows. Optional independent attestation is a separate proposed stage, subject to provider integration and eligibility checks.
The explorer currently exports a fictional JSON example. It does not create a sealed artifact, certify an execution, or connect to a runtime.