Skip to main content

ADR-004: 16 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 16 members after iterating from an initial 9.
The 7 additions (The Oracle, The Envoy, The Sentinel, The Warden, The Steward, The Strategist, The Operator)
filled gaps discovered during real-world use.

Decision

The Society uses 16 specialized members, each with a named role, a defined
trigger condition, and a focused skill file. No general-purpose member exists.

Alternatives Considered

OptionProsConsWhy Rejected
Single general agentSimple; no routing requiredContext bloat; the agent must hold all knowledge simultaneously; no clear ownershipDegrades quality as scope grows
Fewer, broader members (e.g., 4–5)Easier to learnBroad members produce diluted guidance; unclear when to invokeBlurs accountability
16 specialized members (chosen)Each member loads only relevant context; clear trigger conditions; easy to add/removeRequires routing logic; adopters must learn which member to invoke
More members (20+)Maximum specializationDiminishing returns; cognitive overhead for adoptersPremature 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