Atlas links AI coding commits back to prompts

Atlas passed 2,643 stars on GitHub for linking every AI coding agent's commit back to the prompt that produced it. That matters for teams running agents unattended, who need to know why an agent made a change without rereading every diff.

Each run produces a checkpoint tying the commit to the session that made it, including the prompts, tool calls, and reasoning behind it. The tool runs Claude Code, Codex, its own agent, and anything from the ACP registry side by side against one codebase.

Every agent on the codebase also shares memory. Run Claude Code on a feature, then hand the repo to Codex for a bug fix. Codex starts with the context of what Claude Code already decided, instead of working it out again.

That's a different failure mode than agents duplicating work or contradicting each other because nobody told the second one what the first one did. It runs locally by default. Sign-in exists only for team sync, not for the checkpoints or the shared memory to work.

The repo doesn't say how the shared memory is scoped when multiple agents run at the same time rather than one after another.