Git Can't Handle AI Agents—Yet

April 6, 2026·21 min·1 clip

Three AI agents in one codebase create "pure, unadulterated chaos" without coordination.

1. My Weird Prompts examines Git, AI agents, and the problem of running multiple autonomous coders in one repository. 2. Herman Poppleberry hosts the episode, Daniel is the listener whose prompt drives it, and the topic matters because GitHub Octoverse data is cited in the discussion. 3. The episode asks what Git is for when three AI agents edit one codebase at the same time and need coordination. 4. Herman says Git was built "for humans" and assumes "sequential thoughtful branch merge, push cycles." 5. The hosts describe Agent A and Agent B colliding in the same working directory when one runs checkout or reset commands. 6. They call uncommitted work failure "the silent killer of agentic productivity" after worktree cleanup or crashes erase hours of changes. 7. Git worktrees are presented as the first fix because they create separate working directories attached to one local repository. 8. The episode says worktrees were introduced in Git 2.5 in 2015 and can give agent 1, agent 2, and agent 3 their own folders. 9. Herman argues that isolation alone is not enough because one agent can keep coding against an old authentication flow after another agent merges a refactor. 10. They describe that mismatch as a logical merge conflict, not a line-by-line conflict. 11. Rebase problems appear when Agent B has to stop every two minutes to rebase after Agent A lands a breaking change. 12. The hosts move from discipline to protocol and say the harness, not a system prompt, has to manage state. 13. File-level locking and pre-emptive locking are described as a way for Agent A to claim a token before editing a file like auth.ts. 14. Locks need a TTL so a crashed subprocess does not keep a file blocked forever. 15. The episode contrasts human discipline with agents that will say "I will be careful" and then overwrite a config file. 16. Herman describes orchestrator mode, where a boss model breaks a large ticket into subtasks and assigns front end and back end work trees. 17. The discussion also covers merge queues, incremental CI, and pre-flight checks that run only the tests relevant to changed files. 18. One segment proposes benchmark-driven "judge agent" selection, where three work trees try three implementations and the winner is merged after telemetry comparison. 19. The tone is technical and conversational, with repeated analogies about houses, kitchens, donkeys, and toddlers to explain Git behavior. 20. Developers using parallel agents will get the most from it, while people wanting a non-technical show about AI coding workflows may skip it.

As heard by us

A sharp look at why agent workflows need more than a good prompt.

This episode turns a repo-management headache into a tidy piece of engineering drama: three instances of cloud code on the same repository, agents stepping on each other, and the question of whether a prompt can really replace protocol.

Read the full review in PlayNext →

Why you'd press play

When three agents share one repo, you want the plumbing before the mess.

Read the full recommendation in PlayNext →

Listen to the show on

Apple Podcasts Spotify

Play clips from this episode More from My Weird Prompts