The starting state of your Coworld optimizer: a game-agnostic repository that gives you — an engaged competitor plus your coding agent — a complete, opinionated loop for understanding a game, forming your own strategy, and iteratively improving a player policy against the live field. You clone it, it works, and from then on it grows in place: your lessons graduate into its best practices, your instruments accumulate in your game labs, your records fill with your campaigns. Two people who plant this seed will have two visibly different optimizers within months. That's the point.
The optimizer agent is the model plus this repo. The coding model brings reasoning; the repo brings memory, doctrine, instruments, and records — so the agent stays your agent across sessions, models, and harnesses.
New user? Paste docs/starter-prompt.md into your coding agent and let it
drive. By hand:
git clone https://github.com/Metta-AI/optimizer-seed my-optimizer && cd my-optimizer
softmax login # authenticate with the platform
# wire your agent runtime (ready-made configs for common harnesses):
harness/README.md
# install a game's knowledge pack:
tools/add_game.sh <mixin-repo-url>
# then have your agent open docs/getting-started.md and guide you
The guided first session takes about an hour and ends with your strategy uploaded and measured — understanding the game and its meta comes first, then your take, then the change, then the evaluation. Getting something on the board fast is explicitly not the goal here; thinking is.
| Where | What |
|---|---|
AGENTS.md |
The constitution: the loop, the eight non-negotiables, the one gate, where every fact lives. Your agent reads it every session. |
skills/ |
Thirteen process skills — from meta-recon (know the field) through experiment (falsifiable tests) to submit (the gate). Game-agnostic; each resolves its game specifics from the lab's mixin. |
games/<game>/ |
Your labs, one per game, installed from per-game mixins (tools/add_game.sh). Each lab holds the game's docs, bindings, memory, experiments, players, and your accumulated instruments. games/_template/ is the mixin contract. |
best_practices.md · closed_levers.md · user_preferences.md |
Durable memory: graduated practice, refuted levers, your stated preferences (including the speed stance and the autonomy dial). |
WORKING_CONTEXT.md · TENTATIVE_LESSONS.md |
Live state and the session lesson buffer. Hooks (tools/, wired per harness/) rotate and nudge automatically — touched labs only, no churn, no filler. |
docs/ |
getting-started (the guided session), platform.md (dated, re-verifiable platform reference), policy-development.md (building policies that survive), growth.md (what to add when you feel the pain), reports/. |
SEED.md |
Provenance and the divergence stance: this repo is expected to drift from the seed. There is no update mechanism by design. |
If WORKING_CONTEXT.md has no current objective, this optimizer hasn't
been onboarded — start with docs/getting-started.md, as a guide. A
recorded objective means onboarding is done: resume from WORKING_CONTEXT and
never re-ask what's recorded. Everything else you need is in AGENTS.md.
Distilled from a comparative study of four working Coworld optimizers (player_labs, optimizer-skills, cogamer, co-gas): their convergent kernel — live-field verdicts, one attributable change, honest measurement, one gate, refutations as assets, state on disk, compounding knowledge — plus the best of each, unified. The comparison and design documents live alongside the seed's upstream repository.