research-loop

Multi-Agent Research Loop

Architecture (from anthropic.com/engineering/multi-agent-research-system)

Lead Agent (Opus) ─── classifies query
    │
    ├── Subagent 1 (Sonnet) ── explores subtopic A
    ├── Subagent 2 (Sonnet) ── explores subtopic B
    └── Subagent N (Sonnet) ── explores subtopic N
    │
    └── Lead synthesizes ── evaluates gaps ── spawns follow-ups

Process

Phase 1: Classify

Determine query complexity:

• Simple: 0 subagents, direct answer (1-3 tool calls)
• Lookup: 1 subagent, single-source (3-10 tool calls)
• Comparison: N subagents per entity (10-15 calls each)
• Deep dive: N subagents per facet (10-15 calls each)
• Survey: N subagents per subtopic (10-15 calls each, up to 10)

Phase 2: Plan

For each subagent, define:

1. Specific objective (one core question)
2. Output format (summary, JSON, code, or comparison table)
3. Tools to use (Read, Grep, WebSearch, WebFetch)
4. Scope boundaries (what NOT to research)

Phase 3: Execute

Fan out all subagents in parallel using the Agent tool. Each subagent:

• Gets a clean context window
• Explores with 10-15 tool calls
• Returns compressed findings (10K→1K tokens)

Phase 4: Synthesize

The lead agent merges all subagent findings into:

• BLUF (Bottom Line Up Front) — direct answer in first paragraph
• Evidence sections — 3-5 focused sections with bold key facts
• Confidence — explicit confidence level and limitations

Phase 5: Evaluate & Improve (Recursive)

Use Haiku to evaluate the synthesis for gaps:

• If gaps found → spawn targeted follow-up agents
• Merge follow-up findings → re-synthesize
• Repeat up to 3 rounds maximum

Scaling Rules

Query Type	Agents	Tool Calls/Agent	Model
Simple	0	0	haiku
Lookup	1	3-10	sonnet
Comparison	2-5	10-15	sonnet
Deep Dive	2-4	10-15	sonnet
Survey	5-10	10-15	sonnet

Output Format

Always use BLUF (Smart Brevity):

• First paragraph answers the question directly
• Bold key facts and figures
• Narrative prose, not bullet lists
• Source attribution inline
• Final paragraph: confidence level and known limitations