re_gent Wants to Be Git for Coding Agents
A team called regent-vcs put out an open-source tool called re_gent on May 4. Made HN at 85 points. The pitch: "We gave agents write access to our codebases. We did not give ourselves git for it."
What it does: every tool call your AI coding agent makes — Edit, Write, Bash — gets recorded with the prompt that triggered it, the conversation context around it, and a hash that lets you blame any line back to the agent reasoning that produced it. Concurrent agent sessions don't conflict. The /compact compression in Claude Code doesn't lose history because re_gent already captured it.
Apache 2.0, written in 7,800 lines of Go, latest release v0.1.2 May 4, 2026. VSCode extension shipped. Homebrew tap shipped. 146 GitHub stars and climbing. Production-quality code at POC-level feature completeness.
This sits at the intersection of two clusters that have been building for weeks. The harness-safety cluster (Mendral, Rosentic, Tilde, AgentTrust) constrains what agents can do at runtime. The harness-engineering category (learn-claude-code, addyosmani/agent-skills, browserbase/skills) teaches how to architect them. re_gent is the audit-and-rollback layer in between — when the agent does the wrong thing, this is the tool that lets you find out exactly which prompt caused it.
The structural read: 2026 is the year coding agents stop being "AI in the IDE" and become "third-party contributors with their own version control needs." Production teams running Claude Code at scale need an audit trail and a way to reconstruct agent reasoning. re_gent is the first OSS swing at that need. Source: https://github.com/regent-vcs/re_gent
← Back to all articles
What it does: every tool call your AI coding agent makes — Edit, Write, Bash — gets recorded with the prompt that triggered it, the conversation context around it, and a hash that lets you blame any line back to the agent reasoning that produced it. Concurrent agent sessions don't conflict. The /compact compression in Claude Code doesn't lose history because re_gent already captured it.
Apache 2.0, written in 7,800 lines of Go, latest release v0.1.2 May 4, 2026. VSCode extension shipped. Homebrew tap shipped. 146 GitHub stars and climbing. Production-quality code at POC-level feature completeness.
This sits at the intersection of two clusters that have been building for weeks. The harness-safety cluster (Mendral, Rosentic, Tilde, AgentTrust) constrains what agents can do at runtime. The harness-engineering category (learn-claude-code, addyosmani/agent-skills, browserbase/skills) teaches how to architect them. re_gent is the audit-and-rollback layer in between — when the agent does the wrong thing, this is the tool that lets you find out exactly which prompt caused it.
The structural read: 2026 is the year coding agents stop being "AI in the IDE" and become "third-party contributors with their own version control needs." Production teams running Claude Code at scale need an audit trail and a way to reconstruct agent reasoning. re_gent is the first OSS swing at that need. Source: https://github.com/regent-vcs/re_gent
Comments