Dashboard/Algolia-Central (Second-Brain)/v2/wiki/decisions/2026-07-01-rationalize-to-3-agents.md
Algolia-Central (Second-Brain)

v2/wiki/decisions/2026-07-01-rationalize-to-3-agents.md

ADR — Rationalize to 3 agents: General / Developer / Marketer-Merchandiser

Date: 2026-07-01 · Status: Accepted (Arijit) · Supersedes-in-part: 2026-06-18-content-source-multi-agent (the 4-specialist roster) and refines 2026-06-28-content-source-routing-spike-verdict

Context

Two evidence passes converged on the same finding: per-content-type specialists give no answer-quality lift over a single generic agent. - Routing spike (2026-06-28): source SCOPING = no lift on a fair baseline (KILL). - Core-thesis A/B (2026-07-01, 33 Qs, rounds=3, repo scripts/setup/enrich/thesis_specialist_vs_generic.mjs): 4 tuned specialists vs 1 generic all-source agent, same index/model — clean 30-Q means spec 8.07 vs gen 8.41 = flat (gen wins 10 / spec 7 / tie 13). Only marketer/blog showed a specialist edge (~+1.3, n=5). The tech specialist's depth doctrine also looked more grounding-gate-prone.

Decision

Keep multi-agent — but NOT for answer quality. Arijit's call: its value is (1) demoing Agent Studio's multi-agent + handoff capability, (2) distinct personalities, (3) distinct knowledge scopes. Four agents is too many → rationalize to three.

Agent Persona Knowledge scope (source facet on AC2_WWW_MULTI_NEURAL) Role
General Concierge / front desk Website, Other Always interfaces the conversation first; answers broad/company/positioning; routes + opportunistically brings in a specialist.
Developer Technical builder / SE Documentation, Developers, Support, Academy API, how-to, code exchange, troubleshooting/FAQ, tutorials. Folds old tech + support + academy.
Marketer / Merchandiser Business value / ecommerce Blog, Customer Stories, Resources Thought leadership, social proof / case studies, ROI, ebooks/webinars, merchandising narrative.

Source clubbing is disjoint and total — all 9 sources assigned, no overlap (General 2 · Developer 4 · Marketer 3).

Handoff model

General leads every conversation (single front door). Technical intent → warm-baton handoff to Developer; business/value/merchandising intent → warm-baton to Marketer. Handoff carrier = the existing warm-baton dossier mechanism.

Consequences

  • Retrieval scoping = native source facetFilters per agent on the ONE shared index (no new indices — one corpus, three scoped views). Must verify a source-scoped tool actually restricts hits.
  • Collapse old ac2-{support,tech,academy,marketer}-neural; add General. Retire the ephemeral ac2-*-ab-* eval agents.
  • Write 3 persona prompts: shared grounding contract + source scope + persona + handoff doctrine + the "answer only from hits; if hits fall short, say so" grounding escape (fixes the depth-doctrine gate-trip). Keep the marketer/blog long-form handling — the one place specialization measured a lift.
  • P2b judge calibration still gates trust-mode scoring; does not block the build.

Evidence

Repo ADR mirror: docs/experiment/2026-07-01-adr-three-agent-rationalization.md. Thesis results: scripts/setup/enrich/thesis_results.json. Memory: project-three-agent-architecture, project-specialist-vs-generic-verdict.