Researcher Agent

Overview

Layer 4 Research on the Mental Model. The Researcher scrapes only the open slots after catalog and public matches — it does not write the Neo4j plan summary. Public Neo4j matches stay with the Assistant. The harness may call research before the Assistant draft (_fill_open_slots); the Assistant may also call ask_researcher if the plan is still short.

Behavior

• Worker on demand. No person-facing turn (Orchestrator-Worker). See Building effective agents.
• Quota fill. Need = open slots toward three. Skip URLs already on the plan, already offered this session, or already in the public graph.
• Hand off as you go. Each accepted read_page links the recipe onto the plan immediately; stop when the quota is filled, then write one directions summary.
• Respectful fetch. search_web returns scrape; read_page only when allowed (letittoastbot, robots.txt, Crawl-delay). Browser/MCP only if the reader was refused and robots still allow.

Research flow

Assistant already has catalog/public matches; Researcher fills the gap. Diagram key: ✦ AI agent · second line = Anthropic workflow · purple = actor · white = service.

---
config:
  theme: default
  themeVariables:
    background: "#ffffff"
    fontSize: "14px"
---
flowchart TB
    assistant["✦ Assistant
Orchestrator-Worker"]
    specialist["✦ Researcher"]
    web["search web and read page
Playwright MCP if needed"]
    live[Link onto plan]
    handoff[Directions handoff]

    assistant -->|need N remaining| specialist
    specialist --> web
    web -->|each accepted page| live
    specialist --> handoff --> assistant

    classDef agent fill:#e8f0ee,stroke:#0f3d36,stroke-width:2px,color:#0d0d0d
    classDef service fill:#fafaf8,stroke:#c8c8c2,stroke-width:1px,color:#0d0d0d
    class assistant,specialist agent
    class web,live,handoff service

Prompt: researcher_agent.md. Invoked via _fill_open_slots and ask_researcher from develop_plan.