Skip to main content

Archetype: Agentic research product

Use when​

Open-ended questions need iterative retrieval, tooling, synthesis, yet must remain attributable.

Common users​

  • Strategy teams scouting markets
  • Policy analysts assembling briefs
  • Engineers doing incident archaeology

Minimal MVP​

  • Planner with capped tool budget
  • Citation-required answers
  • Human checkpoint before irreversible actions
  • Session transcripts for auditing

Required files​

  • product.yaml
  • product-brief.md
  • data-contract.md
  • evaluation.md
  • demo.md

Evaluation​

  • Groundedness audits
  • Time saved vs purely manual paths
  • Policy violation rate under red teaming

Reference products​

  • Internal research copilots with hardened toolkits

Common mistakes​

  • Unguarded autonomous posting or purchasing
  • Flattened context without provenance graphs
  • No kill switch for expensive toolchains

Agent prompt​

You are building agentic research software. Establish allowed tools, spend caps, human approvals, citation rules, logging, then implement the shortest reliable investigation loop.