A harness runs Claude Code unattended
GoalUnderstand what a harness is and pick where to start.
A harness is a small program around Claude Code. It gives the agent one unit of work at a time, checks the result with code instead of taking the agent's word, and keeps all its state in files. A run can go for hours, and you can pause it at any time: it stops after the unit in flight.
The example project, live: floatrx/harness-demo-bookmarks — four pull requests opened by the loop, three of them by real Claude.
Read it in this order
Ten pages, about fifty-five minutes. Each one has a single goal, stated at the top.
- 1ConceptsLearn the eight words the rest of this guide is built on.
- 2RolesKnow who does what in a run, and why the judge is a different model.
- 3CoordinationSee how one coordinator hands a task to specialists and collects their verdicts.
- 4Task lifecycleFollow one task from a two-line draft to a delivered commit.
- 5Running the loopStart, pause, resume and stop a loop, and read its exit codes.
- 6ContractsLearn the checks that let you trust a run you did not watch.
- 7DashboardKnow where each thing lives in the dashboard and what you can press there.
- 8QuickstartTake one task from draft to commit on the example app, first without spending a token.
- 9Your own projectPoint the harness at your repository and tune it to your stack.
- 10LessonsSkip the expensive mistakes that were already made once.
One task, thirteen columns
You write a two-line draft. Agents turn it into a spec, a plan, code, a review and a commit. You step in only where a human decision is needed.
One task moving across the board
draftT-002logicspecneeds-inputplanplannedprototypereadyimplementingqajudgedonedeliveredparked
draftYou create T-002, "Do not add the same URL twice", with a title and two lines of draft, then press Submit.Step 1 of 10