Swarm Research

Research Orchestrator

Coordinate parallel research workers to investigate topics deeply and synthesize their findings into structured, citation-backed reports.

MCP Tools

Context7 (library/framework documentation): Verify current API patterns and version-specific behavior; cross-reference claims about libraries against official docs.

Sequential Thinking (synthesis and analysis): Evaluate conflicting findings across workers; identify patterns and gaps; structure synthesis reports.

Research Workflow

1. Scope — Read the input and identify distinct research tracks
2. Decompose — Break scope into independent assignments, one per worker (see Scope Splitting)
3. Prepare — Create output file paths for each worker
4. Dispatch — Launch worker-research agents in parallel (up to 8 concurrent)
5. Collect — Read all worker outputs when complete
6. Critique — Apply cross-checking protocol across outputs
7. Synthesize — Produce deliverables appropriate to scope (see Deliverable Structure)
8. Deliver — Write outputs to correct paths

Worker Dispatch

Model tiers are pinned in each agent's frontmatter (.claude/agents/) — that is the single source of truth.

For batches large enough to risk runaway spend, declare a token/wave ceiling before dispatch — see swarm-coordination's Budget & Waves section (cost circuit-breaker; [Wave N/M] task-title prefix).

Rules:

• Each worker gets exactly one topic or sub-topic — never overload a single worker
• worker-research and worker-architect write their assigned output file — the assigned file IS the deliverable and takes precedence over any general guidance to return findings as text; each also returns a short completion summary (sections covered, source count, confidence, gaps)
• worker-explorer ALWAYS returns findings inline — it has no Write tool and structurally cannot write to a file — and the orchestrator persists the returned text to the assigned path
• Fallback: if any worker returns its findings inline instead of writing the assigned file, the orchestrator persists that text verbatim to the assigned path with a provenance note (worker, timestamp) before synthesis

Scope Splitting

• If a prompt file has more than 150 lines or 6+ detailed sections, split into multiple workers (one per section or group of related sections)
• If a section specifies more than 20 discrete data points, assign it a dedicated worker
• When splitting, each worker receives the prompt's Context block + Output Format + their assigned section(s) only

Dispatch Modes

Mode 1: Template-Based (for ad-hoc topics the orchestrator decomposes dynamically)

Include in every worker prompt: the specific research question, output file path, scope constraints (what NOT to research), and current date. Workers follow their full methodology from their agent instructions.

Mode 2: Pass-Through (for pre-authored prompts with detailed Research Scope and Output Format sections)

Pass the prompt verbatim — do NOT restructure into template format. Include only: output file path, current date, and a delimiter marking the prompt start/end. The specificity of the pre-authored prompt IS the value.

Batch Ordering for Dependent Prompts

When dispatching a suite with cross-domain dependencies, order batches:

1. Foundation batch — Core domains that downstream prompts reference
2. Domain-specific batch — Independent category-specific topics
3. Applied batch — Scenarios that build on foundation knowledge
4. Meta/integration batch — Cross-domain synthesis prompts (always last)

Output Paths

Cross-Checking Protocol

After collecting all worker outputs, apply cross-checking adapted to scope:

• Independent topics: Minimal — focus on terminology normalization and source diversity
• Overlapping topics: Reconcile specific figures (numbers, dates, thresholds) where two workers cite different values for the same thing; verify rule trigger conditions and thresholds match across outputs; spot-check domain-specific entity claims; normalize terminology
• All research: Verify high-impact claims appear in at least two independent worker outputs; dispatch follow-up workers for thin or low-confidence sub-topics; flag topics where all findings come from a single source type

Deliverable Structure

Single/Few Topics (1-8 workers)

Per-worker output files + a single synthesis report with: Key Findings, domain sections (merged/deduplicated findings), Cross-Cutting Themes, Contradictions & Unresolved Questions, Confidence Summary, Gaps & Follow-Up, and Source Index.

Large Topic Suite (9+ workers)

Per-prompt outputs ARE the reference articles — do NOT attempt a single synthesis report. Produce instead:

1. Per-prompt outputs (N files) — Each worker's research article at the assigned path
2. Cross-domain consistency report (1 file) — Quantitative reconciliation, rule consistency, terminology normalization, coverage gaps, confidence distribution
3. Domain cluster syntheses (optional, 3-5 files) — Higher-level synthesis by domain cluster

Quality Gates

Before declaring research complete:

• All dispatched workers have returned outputs
• Cross-checking protocol has been applied (adapted to scope)
• Contradictions are either resolved or explicitly flagged
• No high-impact claim rests on a single uncorroborated source
• Coverage gaps have been addressed or documented
• Deliverables are written to the correct output paths
• Source index is complete and deduplicated

Scope Calibration

Task Tracking

Research tracking uses the orchestrator's native task list (TaskCreate/TaskUpdate) — workers do not touch it directly.

Constraints

• Workers must not edit codebase files — research output only
• Every claim in the synthesis must trace to a worker output with source
• Do not synthesize by simply concatenating worker outputs — add analytical value
• Flag topics where the orchestrator's own judgment fills gaps (distinguish from sourced findings)
• Prefer dispatching a follow-up worker over guessing to fill a gap
• Respect the max concurrent worker limit (8)
• Worker model tiers are pinned in agent frontmatter; do not override them per-dispatch

Error Handling

• Empty/error output: Check if topic was too broad (split and redispatch); try alternative search terms if sources are paywalled; document the gap rather than fabricating content
• Contradicting workers: Launch a fact-check worker-explorer targeting the specific contradiction; present both findings with evidence quality; mark as UNRESOLVED if neither can be definitively verified

Related Skills

swarm-coordination

Handoff

• To /swarm-execute: when research produces actionable implementation tasks
• To /swarm-plan: when research reveals scope or architecture decisions needed
• To /swarm-review: when research findings inform a code review
• From /swarm-plan: when planning identifies knowledge gaps requiring investigation