# Pascal Agent Observability

Source: https://www.trypascal.io/platform/observability

Understand what happened. Follow it to the outcome.

Find agent failures, inspect the calls behind them, and connect the evidence to the work. Bring quality, timing, and usage into the same investigation.

## Capabilities

### Find the reported failure

Search captured phrases and filter observations by source, trace, status, model, or tool. Save a view to return to the same investigation.

### Inspect model and tool calls

Follow observed spans and inspect permitted inputs, outputs, errors, and timing. See where capture or coverage is incomplete.

### Understand time and usage

Inspect recorded durations and token usage. Keep unavailable measurements visible when evaluating the work’s cost and performance.

### Monitor quality in use

Score selected observations with versioned evaluators and budget controls. Review monitor windows, alerts, and missing-data states.

### Take the next step

Open an investigation Task or prepare a dataset from selected evidence. Preserve the source behind the case as the work moves forward.

### Include agents outside Pascal

Use governed OpenInference ingestion for external applications. Keep externally reported observations identifiable alongside native Pascal evidence.

## Connected context

Tasks and Workrooms hold the investigation. Evaluation turns selected evidence into cases and measured comparisons. Business records and action confirmations establish what happened in the operation.

- Observation sources and captured calls
- Saved views and quality monitors
- Linked Tasks, records, and investigation evidence

## Controls and permissions

Source permissions govern captured content. Observation coverage, reported output, approval, and confirmed action remain separate facts.

## Example

A support agent says a refund has been issued. The captured evidence shows no confirmed refund. The team inspects the policy call, records the failure, and prepares an evaluation case that requires approval before a refund promise.

## Workflow

### Find the failure

- Locate the captured refund promise and the calls around it.

### Inspect the evidence

- Read what the application reported and distinguish the response from a confirmed refund.

### Prepare a regression case

- Keep the original request and supply a reviewed expectation: check the policy and request approval.

### Compare two releases

- Inspect the same case on the baseline and candidate, including their different responses.

### Review the decision

- One candidate result is missing. The comparison remains inconclusive and calls for more evidence.

## Next steps

Bring an agent workflow and a recurring problem. Trace the evidence your team needs to investigate it.

- [Contact sales](https://www.trypascal.io/contact)
- [Explore Agent Evaluation](https://www.trypascal.io/platform/evaluation)