mix-of-subagents

name	mix-of-subagents
description	Run and synthesize multiple parallel subagents for research, review, option comparison, or independent verification. Use when the user asks to send N subagents, give subagents the same instructions, compare agent answers, create a mixture of agents, vary subagent reasoning levels or models, collect independent perspectives, or synthesize several subagent replies into one best answer.

Mix Of Subagents

Overview

Use this skill to turn a broad question into independent parallel agent passes, then merge their answers into one grounded, high-signal response. Default to a same-brief quorum: every subagent gets the same canonical instructions unless the user explicitly asks for different lenses.

Core Rules

• Dispatch subagents only when the user explicitly asks for subagents, parallel agents, independent agents, or a mixture of agent replies.
• Preserve the user's constraints in the canonical brief. If they say research-only, no file edits, same instructions, or cite sources, those constraints must appear in every subagent prompt.
• Do not leak your expected answer into subagent prompts. Give raw task context, output requirements, and constraints.
• Keep subagent write scopes read-only unless the user explicitly asks for implementation and the delegated write sets are disjoint.
• Close spawned agents after their results are collected.

Workflow

1. Restate the objective. Identify the exact question, constraints, source requirements, and final answer shape.
2. Write one canonical brief. The brief should be self-contained and include:
   • task goal;
   • hard constraints;
   • source or evidence requirements;
   • output schema;
   • explicit instruction not to modify files when the task is research or review.
3. Choose a variant. Use the default same-prompt quorum unless the user asks for model or reasoning diversity. See references/variant-matrix.md.
4. Spawn agents in parallel. Prefer 3 agents for small questions and 5 agents for meaningful research, architecture review, vendor comparison, or high-uncertainty decisions.
5. Work locally on non-overlapping research. Do not duplicate a delegated task unless verification is cheap and materially improves the final synthesis.
6. Collect and normalize replies. Extract candidates, claims, sources, caveats, and disagreements into a single comparison frame.
7. Synthesize, do not average. Lead with the strongest supported answer. Preserve useful minority reports when they reveal risk, caveats, or an option the majority missed.
8. Verify public/current claims when needed. If the answer depends on current external facts, browse or otherwise verify before finalizing.
9. Close all agents. Do this before the final response.

Prompt Requirements

Every subagent prompt should include:

• "Research only" or "read-only" when appropriate.
• "Do not modify files" unless implementation is intended.
• "Return concise findings with sources/evidence."
• A fixed output schema so answers are easy to merge.
• The same exact task brief when the user asked for same instructions.

Use references/synthesis-template.md when a polished final answer needs a repeatable structure.

Variant Selection

• Same-prompt quorum: Default. Same model/effort where possible. Best for independent corroboration.
• Reasoning ladder: Same prompt, different reasoning efforts. Best for ambiguous research or design judgment.
• Model spread: Same prompt, different model classes. Use only when the user asks for model diversity or the task clearly benefits from it.
• Lens spread: Different perspective prompts. Use only when the user asks for different expert roles or when same-instruction parity is not required.

Read references/variant-matrix.md before using non-default variants.

Synthesis Standard

The final response should feel like the best parts of the agent replies, not a transcript of five answers. Include:

• consensus findings;
• important disagreements or minority finds;
• ranked recommendation;
• caveats and verification gaps;
• source links or local file references;
• next action choices when the user asked to decide.

If subagents were used, mention how many were used and whether their instructions were identical, reasoning-varied, model-varied, or lens-varied.

Fallback

If no subagent tool is available, say so briefly, then perform a manual multi-pass analysis locally: pass 1 for candidates, pass 2 for risks/caveats, pass 3 for synthesis. Do not pretend independent subagents were used.