Parallel Sub-Agents — Universal

Purpose

Split a large task into sub-agent tasks that run in parallel without conflicting or duplicating work. Each agent does exactly one job; no two agents share write ownership of the same artifact; no two agents re-derive the same decision.

When to use

• The task can be divided into several independent pieces (multiple files, searches, services).
• The work is large enough that one agent would serially take too long.
• Launching several agents or background tasks at once.

Preconditions

• Delegation tools are available and the current instructions authorize delegation. If not, follow the same partitioned workflow sequentially; do not stop solely because subagents are unavailable.
• Goal, constraints and acceptance criteria are known; use task-intake only if material context is missing, not as a mandatory extra stage for a clear task.
• The interfaces between pieces (signatures, data shapes, file paths) are defined before splitting.

Workflow

1. Decompose by artifact. List the outputs (files, answers, artifacts). Assign one agent per output.
2. Check for interaction. For each pair of agents ask: would both decide something about the same file? would both write to the same path? If yes — merge them, or give one ownership and the other read-only access.
3. Resolve shared unknowns once. Any fact or decision two agents would both need (a design choice, an API shape, a convention) is settled now, before launching, and embedded in every affected prompt.
4. Define contracts first. If outputs must connect (a function and its caller), fix the contract signature before launch — agents then never coordinate mid-flight.
5. Write self-contained prompts. Each prompt carries: the deliverable, the source facts (file paths, exact requirements), the constraints (what not to touch), the acceptance criteria, and the verification. A sub-agent must be able to complete from its prompt alone.
6. Launch within available concurrency limits. Batch scopes rather than exceeding the client's slots. Merge results only where the contracts say they connect; keep confidential work on approved local surfaces.
7. Verify per artifact. Each sub-agent verifies its own deliverable; cross-check integration points yourself.

Rules

• No two agents may own write access to the same file. Reading overlap is fine; writing or deciding overlap is not.
• Give each agent its own search scope (directories, keywords, questions) so explorations do not duplicate.
• Never launch a sub-agent whose prompt ends in a question — it must complete from its prompt alone.
• Merging and conflict resolution belong to the main agent, never to the sub-agents.

Common mistakes

• Two agents fixing the "same" bug in different files because the duplication was not detected before splitting.
• Parallel exploration answering the same question twice with two different results.
• Overlapping write ownership — last-writer-wins silently loses work.
• Sub-agent prompts that require mid-task coordination ("wait until agent B finishes X").

Validation

• The combined result contains each artifact exactly once, written by exactly one agent.
• No piece of work was done twice — run the duplicate check across the merged outputs.
• Each sub-agent completed without asking the main agent a question.