parallel-work

Parallel work

Use the existing authorization and the runtime's available agent tools. This skill organizes delegation; it does not grant permission to spawn agents, spend beyond a budget, publish, or change external systems. A domain skill still supplies the actual method and acceptance criteria. Do not import document-specific gates into coding, research, or ordinary questions.

Choose useful boundaries

Delegate a unit with a clear result that can proceed without another worker's unfinished decisions. For questions, split distinct evidence needs or hypotheses; for production, split independent modules, data partitions, or artifact units. Avoid giving several workers the same broad task unless an independent comparison is intentional. Small tasks or tightly coupled changes can be faster in one agent.

Make dependencies explicit and dispatch ready work. Shared interfaces, terminology, units, and output formats should be settled before their dependent work. Do not finish all target units in the parent before delegating. Batch deterministic work with an existing program rather than assigning an LLM to each mechanical item. Select unit size to amortize context and setup, not one agent per file by default.

Concurrency depends on available slots, model/API capacity, memory, and constrained services. Worker slots and renderer/build/database slots are separate resources. Do not introduce arbitrary iteration caps to hit a speed target. Honor explicit budgets and host limits; unlimited iterations are not permission to run forever.

Establish the common method once

Reuse the applicable tools, domain skills, accepted interfaces, and relevant validation evidence. Verify missing or changed shared capabilities with a small representative check. Every worker should receive usable paths and tool access, not just the instruction to "follow what the parent did."

Share the method, runtime, and legitimate constraints. Let workers solve their assigned data and defects. Do not ask each worker to rebuild the same parser, renderer, test harness, research procedure, or dependency setup. If the task is itself to compare methods, deliberately allow separate implementations and label the experiment accordingly. A common tool failure goes to one owner; update its version and notify affected workers rather than having each invent a workaround.