cluster-agent-setup

Split the initiative into stacks defined by the files each writes, assign every contested file exactly one owner, and give each worker a single cell of a measure/change/verify/publish matrix. One overagent holds priorities and merges while two lead agents own non-overlapping stacks.

Use the RCTF format: role, verified context with provenance (KNOWN, NEIGHBOURS, BOUNDARIES), one concrete task with a stated falsifier, and three to six key results with numeric thresholds written before work starts. A brief missing any of these returns an unverifiable answer.

Do not use it when one agent can finish the work in a single session. Spawning agents costs real budget, so the cluster should be sized to the work, and small one-off tasks gain nothing from roles, pools, and continuity planning.

Write state to disk as it is produced, record outcomes in the task record immediately, and relaunch with a resumption brief carrying the predecessor's measurements inline. Splitting the overagent and leads across separate limit pools prevents total standstill.

Common causes are numbers measured in the wrong tree, two variables changed at once, decorative checks that never fail, and watchers matching only success markers. The failure catalogue lists these modes with guards, and every report must include provenance and a what-I-did-not-do section.