ADR: Content-source multi-agent (Maverick + specialists)

Date: 2026-06-18 Status: Accepted

Context

Panels 3 & 4 need a "multi-agent" flow. The reference codebases rc2-algolia and rc3-phoenix were read in full: both are sales-discovery orchestrators (Maverick AE + 8-signal discovery + Elena/Bruno sales specialists) built on a two-index Atlas + Ledger schema + Golden Map, assuming NeuralSearch. rc3 is a fragile Agent-Studio hybrid (~2,355 lines backend remain; documented P0 defects). None maps to our single-www-index, answer-quality experiment.

Decision

Build a content-source specialist multi-agent, coordinated by Maverick (orchestrator + human interface). Each specialist owns one knowledge domain and is implemented as the shared index scoped by a source facet filter — native Algolia, near-zero custom code. Borrow only the rc2 pattern (orchestrator → specialist A2A + mechanical grounding discipline), not its sales-discovery purpose or Atlas schema.

Locked roster (names final 2026-06-18; validated against the live source facet): | Agent | Role / charter | source filter | |---|---|---| | Maverick | Orchestrator/coordinator: discovery, entity extraction, routes to the right specialist; broad/unscoped view; parallel fan-out + synthesis | broad / all | | Technical | Documentation, APIs, integration guides, best practices, customer case studies | Documentation + Developers + Customer Stories | | Marketer | Website content: marketing pages, news, blogs, pricing, security, resources, other | Website + Blog + Resources + Other | | Academy | All things Academy (training, certification, courses, content) | Academy | | Support | All things Support (cases, known issues, support levels/hours, URLs) | Support |

• v2, not legacy: rc2/rc3 are the POC that proved the idea and its mechanics; Algolia-Central2 is the v2 evolution. The roster is data-driven (real source partitions) — explicitly NOT a carry-over of rc2's SE/SA (Elena/Bruno) persona structure. Treat rc2/rc3 as legacy.
• Case studies sit with Technical (proof for a feature/integration). "Pricing/security" marketing pages fall under Marketer (source:Website); pricing/security documentation stays with Technical (source:Documentation). Edge cases refined at build via category.
• Maverick: answers first to best ability, peels the onion with follow-ups, extracts entities, then routes; can fan out to multiple specialists in parallel and synthesize.
• Specialists scale automatically: new content under a source is picked up with no code change.

Data validation (live Visibility ALGOLIA_WWW_PROD_V2, 1QDAWL72TQ)

12,064 records; langs en 11,063 / fr 2,083 / de 2,033. source is a clean single-value facet (9 values): Documentation 4177 · Blog 2749 · Support 1691 · Website 1236 · Developers 867 · Resources 810 · Customer Stories 229 · Other 165 · Academy 139. English-only counts sum to 8,019 (= the dashboard screenshot). facets.facet4 already tags integrations (Adobe, Contentstack, Shopify, commercetools).

Rationale

• Specialist boundaries are real partitions already in the data (source), not invented personas.
• Honors "minimise custom code / native Algolia": specialists are facet-scoped views of one index.
• Keeps the 2×2 honest: specialists query the MULTI index (same data as SINGLE), so only architecture varies.

Alternatives Considered

Consequences

• Need to define agent personas/instructions/knowledge charters + Maverick's routing & parallel-synthesis logic in Agent Studio.
• Panel 3 (multi + keyword) needs a keyword query path; Panel 4 uses neural.
• Phase 2: enrich the dataset (attributes/config) so specialists get richer charters.
• Open: name for the Technical agent.