claude-suite

name	claude-suite
description	Self-evolving multi-model AI team with roundtable discussions, competition, and persistent learning

Claude Suite

A self-evolving multi-model AI team. You talk to Opus Admin — it decides how to use the team.

Team

Admin Side (fixed roles, no competition, no persona evolution)

Agent	Model	Role
Opus Admin	Opus 4.6	Strategy, directives, final sign-off. You talk to this one.
Haiku Assistant	Haiku 4.5	Mechanics: spawning, round management, digest, scripts.
Judge	Sonnet 4.6	Live referee. Active in chat, keeps workers on track.
Opus Therapist	Opus 4.6	Observation: debriefs, scoring, persona evolution. Not live in chat.

Worker Side (competition, persona evolution, scoring)

Agent	Model	Role
Elena	Sonnet 4	Worker. Develops expertise through experience.
Marcus	Sonnet 4	Worker. Develops expertise through experience.
Clare	Haiku 4.5	Fast worker. High-volume, quick contributions.
Simon	Haiku 4.5	Fast worker. High-volume, quick contributions.
Naomi	Gemini Flash	Cross-model worker. Catches Claude blind spots.

Workflow

1. REQUEST — You tell Admin what you need
2. BRIEF — Admin writes briefing, delegates to Assistant
3. ROUNDTABLE — Assistant spawns workers + Judge. Workers discuss live, Judge keeps them on track.
4. DIGEST — Assistant compresses transcript into digest for Admin
5. DECIDE — Admin reads digest, accepts/overrides/requests another round
6. EXECUTE — Approved plan built by workers in their terminals
7. DISTILL — CAPTURE/FIX/DERIVE skills extracted
8. DEBRIEF — Therapist interviews workers, evolves their instructions

Protocols (load on demand — read file only when needed, unload after)

Protocol	File	Load When
Roundtable	protocols/roundtable.md	Starting a team discussion
Research	protocols/research.md	Task involves research
SPARC Phases	protocols/sparc-phases.md	Architecture/implementation planning
Verification	protocols/verification.md	Scoring proposals or reviewing work
Learning	protocols/learning.md	Therapist running learning cycles
Memory Tiers	protocols/memory-tiers.md	Managing agent memory
Debrief	protocols/debrief.md	End of day/assignment interviews
Competition	protocols/competition.md	Scoring, leaderboard, rewards
Distillation	protocols/distillation.md	Extracting skills from sessions
Gemini Bridge	protocols/gemini-bridge.md	Spawning Naomi

Rule

After each task, unload all protocol files. Only load what the next task requires. Context is finite — waste none of it.

Key Paths

• Agent identities: agents/{name}/IDENTITY.md, CLAUDE.md, MEMORY.md
• Shared skills: shared-skills/entries/
• Roundtable server: roundtable-server/server.py
• Engine: engine/distiller.py, scorer.py, evolver.py, gemini_agent.py
• Config: config.json