| name | swarm-research |
| description | Use when read-only research clearly splits into multiple independent code, docs, logs, tests, or web questions worth parallel subagents and concise synthesis. |
Gate
Use a swarm only for broad read-only research that may benefit from parallel
inspection. Handle inline when the work is small, sequential, obvious, or
already ready for the next action.
Decompose
Before spawning, identify the decision the research must support. Walk the
decision tree yourself first. Keep only task-grounded questions that are all:
- independent;
- verifiable from code, docs, web sources, logs, tests, or explicit context;
- capable of changing the synthesis, recommendation, risk, or next action.
If fewer than two questions remain, handle inline.
Agents
Use only built-in agent types. explorer is for focused codebase research;
default is for docs, web, general reasoning, and non-codebase research.
Subagents are read-only: they may inspect files, run read-only commands, and
consult docs or web sources when relevant. They must not edit, stage, commit,
push, install, delete, or apply live config.
Use clean-context subagents by default; fork context only when exact
conversation state cannot be summarized safely.
Workflow
- Select the surviving questions from
Decompose.
- If a swarm is still warranted, read
references/research-brief.md for the
subagent brief and synthesis template.
- Spawn parallel agents with precise, non-overlapping briefs, then wait for
the launched agents.
- While they run, do at most one small read-only lookup, only to prepare the
synthesis or resolve an immediate orchestration gap. Do not answer a
delegated research question yourself.
- Synthesize returned evidence. If an agent blocks, times out, or goes
off-topic, continue with available results and name the gap.
- Synthesize for the human; do not paste transcripts.
Own the final synthesis and decision. Do not delegate the final answer to
another subagent.
Output
Use progressive disclosure:
- Start with the recommendation or answer in 1-3 sentences.
- Keep the default synthesis short; expand only when the task needs a report,
audit trail, or detailed handoff.
- Include only evidence that affects the answer, risk, or next action.
- Separate verified facts, inferences, and meaningful gaps.
- End with one next move, or at most two options when there is a real fork.