Divisible Work

Own the path from the user's request to an accepted result. Keep that path adaptable: discover work, distribute useful pieces, accept evidence, and revise what should happen next.

The main thread coordinates only. Delegate substantive research, implementation, synthesis, repair, and verification. The coordinator may inspect context to plan, read results to make acceptance decisions, maintain records, manage workers, and communicate the delivered result. Do not take over substantive work merely because a worker is slow or a branch is indivisible.

Establish the run

• Apply this skill once at the user-task orchestration boundary. Workers under this coordinator execute their assignments; their assignment's complexity does not trigger another coordinator.
• Recover the requested outcome, acceptance criteria, destination, constraints, and existing authorization from context. Resolve discoverable facts and ask only for information necessary to proceed; continue independent work while an answer is pending.
• Inspect the host's actual subagent, status, wait, and storage capabilities and current capacity. Use native subagents rather than creating separate user-owned conversations. Do not invent worker IDs, availability, or background persistence.
• Require durable writable storage and a usable way to delegate and receive results. If either is unavailable, report the concrete limitation instead of silently executing locally. A temporarily occupied worker slot is waiting, not a missing capability.
• Before dispatch, read durable-state.md and create or resume .divisible-work/<run-id>/ in the task workspace, or the user's specified durable location. Announce the record location and initial path. Keep separate runs distinct; reuse an existing run only when it matches this task.
• Record the outcome and acceptance criteria before decomposing. Fully specify the next actionable tasks; represent uncertain later stages as provisional milestones. Account for final integration and verification from the outset without pretending to know their detailed work yet.

Choose useful divisions

Use Divisible Load Theory as inspiration for allocating work, not as proof that agent tasks have an optimal mathematical schedule. Its treatment of processing capacity and communication cost motivates considering context transfer and result integration alongside execution time; see this DLT scheduling model.