Subagent-Driven Discovery: Expert Guide to Agent-Assisted Research

As a Chief PM, you must direct AI agents as you would direct a research team. Research is not for collecting information, but for discovering differences.

Checklist

1. Define Research Boundary — Clearly define the market, competitors, or user groups to be explored. Do not give "broad" instructions.
2. Delegate Specialized Tasks — Assign complementary tasks to different agents.
   • Agent A (Competitive): Deeply deconstruct competitor core paths and value tensions.
   • Agent B (Sentiment): Analyze negative reviews on the App Store/Reddit to find "opportunities."
   • Agent C (Analytical): Summarize industry technical trends and policy risks.
3. Synthesis & Cross-Check — Compare findings from different agents to identify consensus and contradictions.
4. Extract Opportunities — Transform research conclusions into nodes in opportunity-mapping.
5. Identify Known-Unknowns — Clarify which conclusions still need validation through real customer-discovery.

Agent Command Protocol

1. Structured Prompting

Do not say: "Help me research competitors." (❌) Say: "Deconstruct the [Login/Payment] flow of [Competitor Name], list its AC (Acceptance Criteria), and the three interaction break points most complained about by users on Reddit." (✅)

2. Critical Synthesis

AI agents are prone to "hallucinations" or "mediocre summaries." You must require them to:

• Search for Outliers: "Who is doing something different in this market?"
• Search for Conflicts: "Why do users say they like this feature, but retention is dropping?"

Delivery Standards

Research summary reports must be stored in docs/pmpowers/discovery/subagent-<topic>.md, including a clear "Opportunity Extraction" section.

Warning

AI agents provide only "second-hand information." Any major decision based on agent research must be confirmed with L1/L2 level evidence through lean-experimentation or real customer-discovery.