Skip to content

ADR-004: 14 Specialized Members Over a General-Purpose Agent

Date: 2026-06-02 Status: Accepted

Context

An AI coding assistant framework could be designed as a single general-purpose agent ("do everything") or as a roster of specialists each responsible for a defined domain. The number and boundaries of specialists needed to be chosen.

The Society settled on 14 members after iterating from an initial 9. The 5 additions (The Oracle, The Envoy, The Sentinel, The Warden, The Steward) filled gaps discovered during real-world use.

Decision

The Society uses 14 specialized members, each with a named role, a defined trigger condition, and a focused skill file. No general-purpose member exists.

Alternatives Considered

Option Pros Cons Why Rejected
Single general agent Simple; no routing required Context bloat; the agent must hold all knowledge simultaneously; no clear ownership Degrades quality as scope grows
Fewer, broader members (e.g., 4–5) Easier to learn Broad members produce diluted guidance; unclear when to invoke Blurs accountability
14 specialized members (chosen) Each member loads only relevant context; clear trigger conditions; easy to add/remove Requires routing logic; adopters must learn which member to invoke —
More members (20+) Maximum specialization Diminishing returns; cognitive overhead for adopters Premature for current scope

Consequences

Easier: Each member's skill file is small and focused — it loads quickly and fits within context budgets. Adding a new member does not affect existing members.

Harder: Adopters must understand enough about the roster to invoke the right member. The Steward exists specifically to solve this routing problem.

New risk: Gaps between member domains can cause tasks to fall through the cracks. The Oracle covers institutional knowledge that does not belong to any specialist.

References