Research Fleet

Dispatch 5–7 subagents simultaneously, each exploring a different research dimension with structured prompts. Combine outputs into a single aggregated deliverable via an executive summary agent.

When to Use

• Broad exploration across multiple dimensions (market research, architecture options, competitive analysis)
• Each dimension is independent — no cross-agent coordination needed during research
• Time-sensitive: 5 agents running in parallel complete in ~5 minutes wall-clock regardless of breadth

Architecture

┌─────────────┐
│  Dispatcher  │
│  (main)      │
└──────┬──────┘
       │ spawns 5-7 parallel subagents
       ├──▶ Agent 1: dimension A → writes /tmp/research-a.md
       ├──▶ Agent 2: dimension B → writes /tmp/research-b.md
       ├──▶ Agent 3: dimension C → writes /tmp/research-c.md
       ├──▶ Agent 4: dimension D → writes /tmp/research-d.md
       └──▶ Agent 5: dimension E → writes /tmp/research-e.md
                                          │
                                          ▼
                                  ┌──────────────┐
                                  │   Aggregator  │
                                  │   (opus)      │
                                  │   reads all   │
                                  │   → gist/doc  │
                                  └──────────────┘

Key Design Points

• Each agent writes to a known file path. Use /tmp/research-<dimension>.md naming. The aggregator reads all files after all agents complete.
• Structured prompts with clear scope. Each agent gets a focused brief: "Research X. Cover these specific aspects. Write findings to /tmp/research-x.md."
• Use a fast, capable model for research agents. Research is read-heavy with web search and synthesis — a model that's fast and good at summarization is ideal.
• The aggregator is the quality gate. Individual research agents may produce uneven quality. The aggregator (usually opus) synthesizes, identifies gaps, and produces the final deliverable.
• Output as a gist for sharing. Multi-file research output is too large for chat. Combine into a single gist with one file per dimension plus an executive summary.

Cost Profile

• Burst cost: all agents run simultaneously, ~3–4 minutes each
• Total wall-clock: ~5 minutes regardless of how many dimensions
• Per-agent cost depends on model choice — cheaper models for data gathering, expensive model only for aggregation

Failure Modes

• Agent scope creep — without tight prompts, agents overlap and produce redundant content
• Silent failure — an agent fails or writes nothing; aggregator proceeds without that dimension. Check file existence before aggregation.
• Web search unavailable — some models or spawn configurations don't have web search. Verify tool availability in the agent config.