First loop

Tell the agent how the loop should run.

After install, assign roles. Init seeds the agent. /arthur-loop is the interview. Then arthur or Run next calls the next role.

1. Create an instance

mkdir ~/my-loop && cd ~/my-loop
arthur init

Init asks which detected CLI is the main agent and which preset you want:

PresetShape
guidedDefault. Manual adapters as a working base; your main agent finishes the setup.
soloOne shipped agent plans and implements (Codex, Claude Code, or Grok Build).
pairMain agent implements; a second shipped agent reviews.
browser-advisorA browser AI (for example ChatGPT Pro) plans and reviews; your main agent implements.
customPick each adapter yourself.

Init writes agent-setup/KICKOFF.md and the skill. Restart the coding agent so MCP is connected.

2. Run /arthur-loop

In Claude Code, Codex, Cursor, or Grok Build, invoke the installed skill or slash command. It asks, one topic at a time:

  • Project id (SHOUTY_SNAKE) and a one-line goal
  • Planner — scopes the packet. Optional model, e.g. grok
  • Implementer — plans and builds. e.g. claude-code
  • Reviewer — plan and sprint review. e.g. claude-code:opus
  • QA — optional. none skips that hop
  • Tracker — atlas-tasker, command, or none

Then it creates the loop with MCP arthur.loop.create. Arthur formulates the default hop sequence. Drive it with arthur, arthur run, or arthur.follow.run.

arthur roles set reviewer=claude-code:opus
arthur loop create --project-id MY_APP --from-ticket AUTH-2 \
  --role planner=grok --role implementer=claude-code --role reviewer=claude-code:opus
arthur
arthur run

Same path in the console: arthur web → Roles, + Loop, Run next. Setup detail: Roles and run. If the tracker is Atlas Tasker and tickets are ready, arthur tracker walk opens them as queue jobs. in_review is not ready work.

Coding agent running /arthur-loop and arthur.loop.create for AUTH_SCOPE
What the installed skill does: interview, then arthur.loop.create.

3. What stays human

Follow will not skip an open human decision, invent an APPROVE_PLAN when the implementation gate is NO-GO, or headlessly drive ChatGPT-browser / manual inboxes. Those pause the project and surface on arthur status and the Needs you rail.

Look first, if you want

arthur init --yes --demo
arthur status
arthur web

The demo is a busy dashboard, not a working agent loop. Use it to learn the map, then run the interview on a real instance.