Your AI Agent Changed the Code — Now You Can Track Exactly Why

Run an AI agent overnight and you'll recognize this situation. Dozens of files have been modified, one test is broken. Which agent decision touched that file? Which prompt triggered that choice? There's no way to know. news.ycombinator

Run git blame and you get a commit message: "fix: update logic." That's it. Why the agent made that decision is nowhere to be found. Git tells you what changed — but not why. re-gent

Why isn't git enough?

Git is doing its job perfectly. The problem is that AI agents produce code in ways git was never designed for.

When a human codes, decisions lead naturally to code, which leads to commits. The "why" lives in commit messages and PR descriptions. But AI agents like Claude Code work differently. A single session involves dozens of prompts, hundreds of tool calls, and modifications to dozens of files — all potentially landing in one commit, or no commit at all. re-gent

And with Claude Code, when the context window fills up, it auto-compacts. Previous messages get summarized and compressed — and any reasoning from earlier in the session is gone forever. re-gent

As AI agents increasingly touch production code directly, if you can't track the decisions an agent made, you're not really in control of your codebase. Bug reproduction, rollback, code review — the foundations of collaboration start to crumble. lyzr

The gap re_gent fills

re_gent doesn't replace git. It's an independent layer that runs alongside git. It automatically captures every agent "turn" between commits as a snapshot. re-gent

The mechanics are similar to git itself. A .regent/ directory appears at your project root, using content-addressed storage plus a SQLite index to record all agent activity. Every time Claude uses a tool — reading a file, editing code, running a terminal command — a snapshot of that moment is saved as a Step. No manual commits, no extra commands needed. re-gent +1

Here are the commands you'll actually use:

• rgt log — full history of agent decisions
• rgt blame [file] — trace which prompt produced a specific line of code producthunt
• rgt sessions — list sessions and their work
• rgt checkout [step] — return to a specific Step in time
• rgt replay — replay what the agent did, step by step

Getting started in 5 minutes

Public alpha launched June 2, 2026. Apache-2.0 open source, completely free. re-gent

1. Install (macOS / Linux)
   macOS: brew tap regent-vcs/tap && brew install regent
   Go users: go install github.com/regent-vcs/re_gent@latest
2. Initialize in your project
   Run rgt init in the root of any project where you're using Claude Code.
   A .regent/ directory appears and recording begins.
3. Run your agent as usual
   Use Claude Code, Codex, or OpenCode normally. re_gent silently records every agent turn in the background.
4. Check the history
   rgt log shows the full decision trail. rgt blame [filename] traces exactly which conversation produced a specific line.
5. Roll back to a specific moment
   rgt checkout [step-id] returns you to that exact point. Think of it as git bisect, but for agent decisions.

Want to dig deeper?

re_gent Official Site Installation guide, supported agents, roadmap re-gent.dev

GitHub — regent-vcs/re_gent Source code, README, issue tracker github.com

Show HN: Git for AI Agents HN community reactions and discussion of competing tools news.ycombinator.com

Version Control for AI Agents: The Missing Layer in Enterprise AI Technical analysis of why agent versioning matters lyzr.ai

What Is AI Agent Observability? A 2026 Guide Full context on agent observability voiceflow.com

Former GitHub CEO raises record $60M dev tool seed round The story behind Entire.io, tackling the same problem from an enterprise angle techcrunch.com