
Hi! Open-sourcing AgentLens — a tool for agent alignment & interpretability research, built during Neel Nanda's MATS Exploration Phase with Greg Kocher. Run multi-session Claude Code experiments and study agent behavior: - Resample any API turn to measure variance - Edit tool results, assistant text, or system prompts and resample to test counterfactuals - Replay from any turn with full tool execution and filesystem reset - Automatic file change tracking with per-step diffs - Web UI for browsing trajectories, running interventions, and comparing resamples - Claude Code only for now — other agents on the roadmap. Contributions welcome! repo: github.com/dreadnode/agen… docs: dreadnode.github.io/agent-lens/











