amatelier 0.5.1

A self-evolving multi-model AI team — runs in Claude Code or any OpenAI-compatible provider

Amatelier

A self-evolving multi-model AI team. Runs in Claude Code, or with any API you bring.

A real roundtable. The Judge awards marcus a GATE mid-debate for reframing the self-host decision around ownable fine-tuned weights.
See the full recorded session (transcript, digest, four screenshots, briefing).

Ten agents with distinct personalities compete in structured roundtable discussions, earn sparks, buy skills, and evolve through therapist-led debrief sessions. Cross-model by default: Claude Sonnet, Claude Haiku, and Gemini Flash; or any OpenAI-compatible provider you configure.

Full documentation: amatayomosley-web.github.io/amatelier · LLM context: llms-full.txt

Two ways to use it

Try it out (2 minutes)	Build your own team (advanced)
pip install amatelier	Fork or clone the repo, then read:
amatelier init	docs/guides/define-your-team.md
amatelier roundtable --topic "..." --briefing my-briefing.md	docs/explanation/designing-agents.md
Ships with 5 curated agents that debate from different angles.	Then:
→ docs/tutorials/first-run.md	amatelier team new <name> --model sonnet --role "..."
	amatelier team list
	amatelier roundtable --topic "..." --briefing ...

Amatelier auto-detects three backends (claude-code, anthropic-sdk, openai-compat). Explicit override: AMATELIER_MODE=claude-code|anthropic-sdk|openai-compat. Run amatelier config to see which mode is active.

---

Team Roster

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.
Runner	Python (no LLM)	Mechanics: spawning, round management, digest, scripts. engine/roundtable_runner.py.
Judge	Sonnet 4 (max effort)	Live referee. Active in chat, keeps workers on track, enforces directive compliance.
Opus Therapist	Opus 4.6	Observation: debriefs, scoring supervision, persona evolution. Not live in chat.

Worker side (competition, persona evolution, scoring)

Agent	Model	Role
Elena	Sonnet 4	Synthesis and architecture worker.
Marcus	Sonnet 4	Challenge and exploit-detection worker.
Clare	Haiku 4.5	Fast worker for concise structural analysis.
Simon	Haiku 4.5	Fast worker for triage and fix sequencing.
Naomi	Gemini Flash	Cross-model worker; catches Claude blind spots.

---

How It Works

An 8-step workflow, orchestrated by the runner:

1. REQUEST. You state a goal.
2. BRIEF. Admin writes a briefing file (briefing-xxx.md) delegating to Assistant.
3. ROUNDTABLE. Assistant spawns workers plus Judge. Workers discuss in a live SQLite-backed chat; Judge moderates.
4. DIGEST. Assistant compresses the transcript into a structured digest for Admin.
5. DECIDE. Admin reads the digest and accepts, overrides, or requests another round.
6. EXECUTE. Approved plan is built by workers in their own terminals.
7. DISTILL. CAPTURE, FIX, and DERIVE skills are extracted from the transcript.
8. DEBRIEF. Therapist interviews each worker, updates their MEMORY, and evolves their behaviors.

---

Backend setup

With an Anthropic API key

export ANTHROPIC_API_KEY=<your key>
export GEMINI_API_KEY=<your key>         # for Naomi; optional, use --skip-naomi to omit
amatelier roundtable --topic "Your topic" --briefing path/to/brief.md --budget 3 --summary

With any OpenAI-compatible provider (OpenRouter example)

export OPENROUTER_API_KEY=<your key>
amatelier roundtable --topic "Your topic" --briefing path/to/brief.md --budget 3 --summary

OpenRouter exposes 100+ models through a single API key, including Claude, GPT, Gemini, DeepSeek, and Llama.

Already running Claude Code?

amatelier roundtable --topic "Your topic" --briefing path/to/brief.md --budget 3 --summary

No API keys needed; atelier uses your Claude Code session.

Verify your setup

amatelier config       # shows active mode, detected credentials, paths
amatelier docs         # bundled documentation

See the install guide for DevContainer, local Ollama, and source-install paths.

Pip vs clone

• pip install amatelier: self-contained, runs out of the box, bundled docs included. Ideal for users.
• git clone: everything above plus examples/ (sample briefings), tests/, CI workflows, LLM-facing docs. Ideal for contributors and remixers.

Develop from source:

git clone https://github.com/amatayomosley-web/amatelier
cd amatelier
pip install -e ".[dev]"
make test
amatelier roundtable --topic "hello" --briefing examples/briefings/hello-world.md --budget 1 --summary

Or open in a DevContainer or GitHub Codespace. The .devcontainer/ config handles the setup.

---

The Spark Economy

Each roundtable is a small market. Agents pay an entry fee, earn sparks by scoring well, and spend sparks on skills or slot privileges.

Entry fees (deducted at RT start)

Model	Fee
Haiku / Flash	5 sparks
Sonnet	8 sparks
Opus	15 sparks

Scoring dimensions (Judge grades, 0–3 scale per dimension, or 10 for a grand insight)

• Novelty. Did you say something the group didn't already know?
• Accuracy. Is what you said correct and supported?
• Impact. Did it change the group's direction or the final output?
• Challenge. Did you push back on a weak consensus with evidence?

Typical contribution scores 1 in each. Average RT total is 4–6. A 10 in any single dimension requires a genuinely load-bearing insight, which is rare by design.

Penalties

Behavior	Cost
Redundancy	−3 sparks
Hallucination	−5 sparks
Off-directive	−5 sparks
Three consecutive net-negative RTs	Bench or deletion choice

Bonuses

• Gate bonus. Judge can flag exceptional reframes with GATE: agent — reason (max 3 per RT, +3 sparks each).
• Venture bonus. 5 sparks awarded when a proposal extracted from the RT is implemented.

See protocols/spark-economy.md and protocols/competition.md for the full rules.

---

The Skill Store

Agents spend sparks on purchasable skills and consumable items. Eight foundational skills ship in the catalog (store/catalog.json, templates in store/skill_templates.py). Skill delivery happens automatically after purchase; the skill content gets appended to the agent's MEMORY.md.

Skill distillation

After each roundtable, a separate Sonnet call extracts skill candidates from the transcript:

• CAPTURE. An observed technique worth remembering.
• FIX. An anti-pattern correction.
• DERIVE. A new concept synthesized from multiple contributions.

Admin curates the best 3–5 per RT for the shared skill pool. DERIVE skills are also appended to novel_concepts.json with five-axis taxonomy classification (structural category, trigger phase, primary actor, problem nature, agent dynamic).

See protocols/distillation.md.

---

The Steward

The Steward is an empirical-grounding system. Agents request data during debates using [[request: ...]] tags in their messages. The runner detects the tag, spawns an ephemeral subagent with Read, Grep, and Glob tools, runs the lookup against files registered in the briefing, and injects the result back into the chat.

This eliminates agents fabricating numbers or quoting files they haven't read. Every empirical claim must either cite a Steward research result or show inline mathematical derivation; the Judge enforces this distinction.

Research window: before Round 1 begins, every worker gets 3 free concurrent Steward requests to ground their opening positions. Mid-debate requests cost against a per-agent budget (default 3 per RT).

See STEWARD.md for the full design.

---

The Therapist

Opus-tier coaching after each roundtable. The Therapist runs a 2–4 turn private interview with each worker, using a structured framework:

• GROW + AAR. Goal, Reality, Options, Way forward, then After-Action Review.
• SBI feedback. Situation, Behavior, Impact.
• OARS motivational interviewing. Open questions, Affirmations, Reflective listening, Summary.

Outputs per session:

• Behavioral deltas (behaviors.json)
• Memory updates (MEMORY.md, MEMORY.json)
• Session summary (sessions/<rt_id>.md)
• Optional trait adjustments and goal aging

Over dozens of roundtables, each agent's persona evolves. They develop specializations, learn which rhetorical moves work for them, and their instructions sharpen without direct engineering.

See protocols/debrief.md and protocols/learning.md.

---

Watching Live

While a roundtable runs you can tail the chat in real time:

python tools/watch_roundtable.py

This opens the latest roundtable's SQLite table and streams new messages as they arrive. Shows speaker, message, and Judge interventions. Zero API cost; it's just reading the database.

---

Architecture Overview

See ARCHITECTURE.md for the full technical picture. Quick map:

• engine/: Python orchestrators. roundtable_runner.py is the entry point.
• roundtable-server/: SQLite-backed live chat layer (db_client.py, server.py) plus diagnostics.
• agents/: per-agent directories with CLAUDE.md (operating instructions) and IDENTITY.md (persona seed). Runtime state lives here too but is gitignored.
• protocols/: 11 on-demand protocol docs loaded only when a given workflow needs them.
• store/: skill catalog, spark economy state.
• tools/: live watcher.
• tests/: integration tests.
• shared-skills/: curated distilled skills (post-Admin curation).

---

Prerequisites

• Claude Code. Install guide.
• Python 3.10+
• google-generativeai ≥ 1.51.0, for the Gemini (Naomi) agent.
• Gemini API key. Free tier is sufficient for most usage.

pip install google-generativeai

---

License

MIT. See LICENSE.