research

What this skill does

# Multi-Agent Research

Wave-based knowledge gathering with deferred synthesis. Mined from 300+ real research dispatches: the pattern that consistently produces actionable intelligence.

**Core insight:** Research breadth-first, synthesize after. Conclusions drawn from the first three results miss nuance the fourth wave would have surfaced. Deploying agents in waves and accumulating findings before synthesizing produces sharper recommendations.

**How to read this skill:** the wave structure below is a shape, not a procedure. Quick mode skips most of it. Standard research uses one wave plus targeted follow-ups. Deep dives genuinely need the full pattern. Calibrate to the question, not the framework.

## The Shape

```dot
digraph research {
    rankdir=TB;
    node [shape=box];

    "1. PRIME" [style=filled, fillcolor="#e8e8ff"];
    "2. WAVE 1: Broad Sweep" [style=filled, fillcolor="#ffe8e8"];
    "3. GAP ANALYSIS" [style=filled, fillcolor="#fff8e0"];
    "4. WAVE 2+: Targeted" [style=filled, fillcolor="#ffe8e8"];
    "5. SYNTHESIZE" [style=filled, fillcolor="#e8ffe8"];
    "6. DECIDE & RECORD" [style=filled, fillcolor="#e8e8ff"];

    "1. PRIME" -> "2. WAVE 1: Broad Sweep";
    "2. WAVE 1: Broad Sweep" -> "3. GAP ANALYSIS";
    "3. GAP ANALYSIS" -> "4. WAVE 2+: Targeted";
    "4. WAVE 2+: Targeted" -> "3. GAP ANALYSIS" [label="still gaps", style=dashed];
    "3. GAP ANALYSIS" -> "5. SYNTHESIZE" [label="coverage sufficient"];
    "5. SYNTHESIZE" -> "6. DECIDE & RECORD";
}
```

---

## Phase 1: PRIME

Lean on existing knowledge before spawning agents. Re-running research that already lives in Sibyl burns tokens and produces duplicate entries.

### Common moves

- **Search Sibyl first:** `sibyl search "<research topic>"`, `sibyl search "<related technology>"`, `sibyl search "<prior decision in this area>"`. Surface what's already known before generating new findings.
- **Check for staleness.** Fast-moving topics (frameworks, models, cloud services) usually warrant re-research even when Sibyl has recent entries; treat the existing knowledge as a baseline. Stable topics with recent entries often don't need a fresh pass at all.
- **Sharpen the research question.** "Research databases" is too vague to dispatch on. "Compare PostgreSQL vs CockroachDB for multi-region write-heavy workloads with <10ms p99 latency" gives agents enough scope to do useful work.
- **Calibrate the research budget** to the decision the research is feeding:

   | Depth          | Agents | Time      | When                                         |
   | -------------- | ------ | --------- | -------------------------------------------- |
   | **Quick scan** | 2-3    | 2-5 min   | Known domain, just need latest info          |
   | **Standard**   | 5-10   | 10-15 min | Technology evaluation, architecture options  |
   | **Deep dive**  | 10-30  | 20-40 min | Greenfield decisions, SOTA analysis          |
   | **Exhaustive** | 30-60+ | 40-90 min | New project inception, competitive landscape |

### Source quality contract

This bit is non-negotiable: the value of research collapses when claims rest on stale blog posts. Specific claim types deserve specific source standards:

| Claim type              | Preferred source                                 |
| ----------------------- | ------------------------------------------------ |
| Current version         | Package registry, release page, or official CLI  |
| CLI flags / config keys | Official docs or local `--help` output           |
| Security frameworks     | OWASP, NIST, SLSA/OpenSSF, CIS, ISO, PCI sources |
| Cloud/provider behavior | Provider docs and current changelog              |
| Research papers / SOTA  | Paper, benchmark repo, or authors' artifact      |
| Community health        | Repository activity plus issue/release cadence   |

When primary sources disagree with secondary ones, trust the primary source and note the discrepancy. Date volatile facts explicitly, and prefer commands/sources the next agent can rerun over screenshots that go stale.

---

## Phase 2: WAVE 1: Broad Sweep

Deploy the first wave of agents across the full research surface. The goal is breadth; accept that some agents will produce mediocre output, that's what gap analysis is for.

### What good agent prompts have

Vague prompts produce vague research. Each agent benefits from:

- **One specific topic** (not "research everything about X")
- **An output file path** (no ambiguity about where to write)
- **Search hints** (include year: "search [topic] 2026")
- **8-12 numbered coverage items** that scope the research precisely
- **Source quality guidance** ("prefer official docs and GitHub repos over blog posts")

### Wave 1 Template

```markdown
Research [SPECIFIC_TOPIC] for [PROJECT/DECISION].

Create a research doc at docs/research/[filename].md covering:

1. Current state (latest version, recent changes)
2. [Specific capability A relevant to our use case]
3. [Specific capability B]
4. [Integration with our stack: list specific technologies]
5. Performance characteristics / benchmarks
6. Known limitations and gotchas
7. Community health (stars, activity, maintenance)
8. Comparison with alternatives (name 2-3 specific alternatives)

Use WebSearch for current information. Include dates on all facts.
Cite sources with URLs.
```

### Deployment notes

- **Background by default.** Research agents have no inter-dependencies, so foreground execution serializes work that should run in parallel.
- **3-4 seconds between dispatches** avoids rate limiting in practice. Tighter cadences sometimes work, sometimes hit limits, so pace yourself.
- **One file per agent.** Shared outputs create write contention and lose attribution.
- **Group by theme** when researching many topics. 12 separate dispatches become 3-4 thematic clusters with clearer synthesis later.

### Coverage Strategy

For technology evaluations, cover these dimensions:

| Dimension       | Question                          |
| --------------- | --------------------------------- |
| **Capability**  | Does it do what we need?          |
| **Performance** | Is it fast enough?                |
| **Ecosystem**   | Does it integrate with our stack? |
| **Maturity**    | Is it production-ready?           |
| **Community**   | Will it be maintained in 2 years? |
| **Cost**        | What does it cost at our scale?   |
| **Migration**   | How hard is it to adopt/abandon?  |

---

## Phase 3: GAP ANALYSIS

After Wave 1, look for what's missing before synthesizing. Premature synthesis is the most common research failure: the answer feels obvious after three docs and turns out to be wrong after eight.

### What to look for

- **Coverage gaps**: dimensions the wave didn't touch, missing comparisons, questions raised but not answered
- **Contradictions**: agents reaching different conclusions on the same question (often signal for verification agents)
- **Bias signals**: all-positive findings (suspicious, look for failure cases), only-official-docs (need community experience), same sources cited repeatedly (need source diversity)

### Decision Point

| Finding                        | Action                                |
| ------------------------------ | ------------------------------------- |
| Good coverage, minor gaps      | Synthesize now, note gaps             |
| Significant gaps               | Deploy Wave 2 targeted agents         |
| Contradictory findings         | Deploy verification agents to resolve |
| Entirely new direction emerged | Deploy Wave 2 in new direction        |

---

## Phase 4: WAVE 2+: Targeted Research

Fill specific gaps identified in the analysis. Wave 2 agents differ from Wave 1 in shape:

- **Smaller scope**: one specific question per agent
- **Higher quality bar**: "find production experience reports, not just docs"
- **Cross-reference prompts**: "Agent X found [claim], verify against [alternative source]"
- **Deep reads**: "Read the full README and API docs for [library], not