agex: running coding agents in parallel

A CLI that gave every agent task its own git worktree, built when agents were far less predictable than they are now.

I built agex in the early days of agentic coding. Agents were extremely non-deterministic back then and hallucinated a lot, so running one straight on your main branch often meant cleaning up after it. Git worktrees solved that by giving each attempt its own copy of the repo, but a fresh worktree doesn't have your node_modules or your .env files, so every new one meant rebuilding dependencies and copying secrets over by hand.

agex created a worktree for each task with node_modules linked to the ones in your main checkout, so no worktree had to rebuild dependencies, and it copied in your .env files and any other secrets that aren't in git. It also gave each task its own port so parallel runs didn't clash, and it let you check each attempt against your tests, compare them side by side and merge the one you liked, without touching main until you accepted it. It worked with Claude Code, Codex CLI and Copilot CLI, and with any agent that can run shell commands.

I've retired it now because a lot of these tools support worktrees out of the box, but the code is still on GitHub and npm.