Build a team of AI coding agents the way you'd build a real one: managers who split the work, workers who do it, reviewers who check it. You give the top agent one goal and a definition of done. Nothing counts as done until the checks pass.
curl -fsSL https://raw.githubusercontent.com/arranger-dev/arranger/main/install.sh | shOne agent with one huge prompt drifts, forgets and claims success. A team with small, checked goals doesn't, because each layer only has to get one thing right and prove it.
You describe the outcome once. A manager splits it into one subgoal per team member, and every subgoal must come with a shell check or the plan is rejected.
Every agent gets its own git worktree and branch. Siblings work at the same time, and nobody overwrites anybody else. Your own branch is never touched.
Arranger runs the checks itself after each attempt. Failures go back to the agent with the exact output, up to three tries, then it stops and tells you.
Managers read each child's diff and accept it or send it back with feedback. Only work that passed both checks and review flows up, so one agent's mistake doesn't spread to the rest.
Each worker sees its own goal, its manager's goal and nothing else. Short, focused prompts are where agents are most reliable.
Live logs and status for every agent, a diff per agent, remove or promote single changes, revert any checkpoint, token limits per agent, and one-click merge into the branch you choose.
Drag agents into a tree. Pick a tool for each one and save the team as a project on your repo.
Give the top agent the outcome, acceptance criteria, and checks like go test ./....
Managers plan, workers build, checks gate, reviewers accept or send work back. Watch it live.
Review the final diff, drop what you don't want, and merge into main or any branch.
Use the best tool for each role, or compare them side by side. Arranger drives them headlessly and reads their output live.
A single Go binary with an embedded UI and a SQLite file. No cloud account, and your code stays local.