Skip to content

Lunette

Lunette captures your agent's trajectories and helps you understand what's happening.

When your AI agent runs—whether it's solving coding tasks, answering questions, or using tools—Lunette records every step. You can then browse these trajectories, see where things went wrong, and launch investigator agents that analyze failures for you.

Quick Start

The fastest path is with Inspect AI:

pip install lunette-sdk
inspect eval your_task.py --sandbox lunette

Your trajectories are now being captured with full environment access.

Getting Started Guide

How It Works

  • Record


    Capture trajectories as your agent runs. Use tracing for any LLM, or run agents in Lunette sandboxes.

  • Investigate


    Launch AI investigators that analyze your trajectories. They can read transcripts, search for patterns, and—with environment mode—execute commands in the original sandbox.

  • Find Issues


    Investigators create structured issues with evidence, confidence scores, and message references. Find test mis-specifications, environment problems, and agent failures.

  • Web App — Browse trajectories and launch investigations
  • Demo — Try the platform
  • GitHub — Source code