Skip to content
Perseval Perseval Docs

Find what keeps failing

Local-first observability for agent systems
Install Perseval Send your first trace

Perseval collects complex agent traces, surfaces recurring failures across runs, and keeps every diagnosis connected to inspectable evidence.

What it helps you do

Section titled “What it helps you do”
  • Find recurring failures: group related failures across runs and rank them by severity, recurrence, affected runs, trend, and recovery.
  • Verify diagnoses: inspect expected and observed behavior, representative examples, evidence spans, telemetry gaps, and provenance.
  • Create eval definitions: turn one or several accepted failure groups into reviewable eval candidates with cases, expected behavior, rubric, grader, and evidence.
  • Explore large traces: navigate virtualized runs and lazy trace hierarchies with agent roles, events, links, attributes, and bounded payload previews.
  • Compare runs: align baseline and candidate execution and move to the first meaningful divergence.

A practical path through the docs

Section titled “A practical path through the docs”
  1. Install the macOS application and create a project.
  2. Send a first trace with the project routing attribute.
  3. Add the behavioral telemetry contract so errors can become evidence-backed findings.
  4. Use the Failure Inbox, Investigation, and Full Trace to verify a recurring failure.
  5. Compare a baseline with a candidate and create a reviewable eval definition.
  6. Connect a coding agent through the bundled, read-only MCP server.

The current limitations page is part of the product contract. It separates shipped behavior from planned eval execution, provider durability, and platform work.