Specialist Interrogators

The Scenario

An idea enters ideation: 'add single sign-on to the customer portal.' The cell could run one large interrogation that asks every question at once - and either skip half of them or exhaust the human's patience asking all of them. Instead it decomposes ideation into specialist interrogators, one per orthogonal concern, each producing a typed decision-brief that downstream agents read selectively. This example illustrates the decomposition pattern, not a mandatory roster.

The Roster (Example)

The number five is not magic. A payments domain might add a regulatory-interrogator; a data product might drop ux-ui. The pattern is one interrogator per orthogonal concern, not this specific list.

A Sample Brief

Each interrogator writes a typed artifact into the idea's folder. The security-interrogator produces, for the SSO idea, a SecurityDecisionBrief:

brief: SecurityDecisionBrief
idea: IDEA-014-portal-sso
produced_by: security-interrogator
findings:
  data_sensitivity: high          # customer identity + session tokens
  authn:
    protocol: OIDC                # decided with human; SAML rejected (no legacy IdP)
    token_storage: http-only-cookie
  authz:
    model: role-based
    blast_radius: organizational  # a flaw exposes every portal account
  compliance:
    - GDPR (EU customers in scope)
    - SOC2 (existing control set applies)
  open_questions:
    - "Is step-up auth required for profile changes, or only at login?"
  trifecta_flags:
    untrusted_content: true       # user-supplied login parameters
    private_data_access: true     # customer records
    external_communication: true  # IdP callback
    note: "Lethal-trifecta candidate - see 10.1; force HITL on the auth callback agent."
confidence: 0.78

The downstream spec-writer reads this brief plus any others in scope; the technical-interrogator's brief feeds dialect selection, the infrastructure-interrogator's feeds the eventual deploy stage.

What This Illustrates

• One Concern, One Interrogator - the questions for 'is this secure?' are genuinely different from 'is this fast enough?'; splitting by concern keeps each interrogator focused and gives the human meaningful pause points
• Typed Briefs Are the Interface - each output is a structured brief with a declared type, not free prose, which lets the spec-writer consume exactly the briefs in scope
• The Roster Is Cell-Level - the decomposition is the contribution; the specific five are an example, which is why the pattern lives in reference examples rather than the normative body
• It Composes With the Rest of the Model - each brief carries a confidence value, each interrogation is recorded in the decision trace, and a brief that raises a lethal-trifecta flag forces the downstream agent to HITL

Each brief carries a confidence value, each interrogation is recorded in the decision trace, and a brief that raises a lethal-trifecta flag forces the downstream agent to HITL. The specialists are ordinary capabilities; the only new idea is splitting ideation by concern. This pattern correlates to the Interrogatory LLM role described by Martin Fowler - the umbrella idea-interrogator orchestrates the orthogonal specialists, each acting as a focused interrogatory agent over its single concern.