research-and-review

Research and Review: Parallel Investigation Pattern

Overview

Multiple teammates investigate different aspects simultaneously, then share findings and challenge each other's conclusions. The key value is adversarial review — not just parallel work, but structured debate.

Core principle: Investigate in parallel, challenge in rounds, synthesize with evidence.

Management theory: Tuckman's Storming phase is intentional here. Psychological Safety (Edmondson) ensures agents flag problems without hesitation.

When to Use

• Problem has multiple angles that can be explored independently
• You need diverse perspectives, not just speed
• Findings should be cross-checked and challenged
• Single investigator might anchor on first plausible explanation

Don't use when:

• Investigation is linear (step A informs step B)
• Only one aspect to research
• Speed matters more than depth

Team Composition

coordinator (lead)
├── researcher × 2-3    (each investigates a different aspect)
└── devil-advocate × 1  (challenges all findings)

Belbin coverage:

• Thinking: researcher (Specialist) + devil-advocate (Monitor-Evaluator)
• Action: researchers also implement their investigation
• People: coordinator (Coordinator) synthesizes

Sizing: 3-4 total (coordinator + 2-3 investigators). Devil-advocate can double as one of the researchers if team is small.

The Process

Phase 1: Forming — Scope & Assign

The coordinator:

1. Defines the research question clearly
2. Splits into independent investigation areas
3. Spawns researchers, each with a distinct scope
4. Spawns devil-advocate with mandate to challenge

Spawn prompt template:

You are investigating [ASPECT] of the problem: [PROBLEM].

Your scope: [SPECIFIC AREA]
Do NOT investigate: [OTHER AREAS — those are covered by teammates]

When done, report:
- What you found (with evidence: file paths, line numbers, logs)
- Your interpretation
- Confidence level (high/medium/low)
- What could disprove your finding

Phase 2: Storming — Independent Investigation

Each researcher works independently:

• Reads code, runs tests, checks logs
• Forms conclusions with evidence
• Reports findings to coordinator

Devil-advocate waits for findings, then:

• Reviews each researcher's evidence
• Identifies gaps, assumptions, contradictions
• Proposes alternative interpretations
• Asks: "What evidence would disprove this?"

Phase 3: Norming — Structured Debate

Coordinator facilitates:

1. Each researcher presents findings
2. Devil-advocate challenges each finding
3. Researchers respond with evidence or acknowledge gaps
4. Coordinator identifies consensus vs. disagreements

Debate protocol:

For each finding:
  1. Researcher states conclusion + evidence
  2. Devil-advocate states objection or alternative
  3. Researcher rebuts or concedes
  4. Coordinator records: CONFIRMED / DISPUTED / NEEDS MORE EVIDENCE

Phase 4: Performing — Resolution

For CONFIRMED findings: include in final synthesis For DISPUTED findings: assign follow-up investigation For NEEDS MORE EVIDENCE: researcher gathers more data

Phase 5: Adjourning — Synthesize & Record

Coordinator produces:

• Final synthesis with confidence levels
• Dissenting views (if any remain)
• Lessons learned → trigger team-orchestrator:session-reflection

Example: Investigating a Performance Regression

Coordinator creates team:
  Researcher A → "Investigate database query changes in last 2 weeks"
  Researcher B → "Investigate frontend bundle size and rendering"
  Researcher C → "Investigate infrastructure/deployment changes"
  Devil-advocate → "Challenge all findings, propose alternative causes"

Results:
  A: "Found N+1 query introduced in commit abc123" (high confidence)
  B: "Bundle size unchanged, no rendering issues" (high confidence)
  C: "No infra changes in timeframe" (high confidence)
  Devil-advocate: "A's finding explains DB load but not the 2s delay
    on pages that don't hit that query. Check CDN cache invalidation."

  → Follow-up investigation reveals CDN config change was the second cause.
  → Without devil-advocate, team would have stopped at the N+1 query fix.

Common Mistakes

Mistake	Fix
Researchers investigate same area	Give explicit non-overlapping scopes
Devil-advocate too agreeable	Prompt must say "finding no issues is a failure"
Skipping debate phase	Debate is the VALUE — don't skip it
Coordinator anchors on first finding	Wait for ALL researchers before synthesizing
No evidence, only opinions	Require file:line references or test output

Integration

Pre-requisite: team-orchestrator:orchestrating-work routes here Post-requisite: team-orchestrator:session-reflection records learnings