World Model Optimizer

wmo optimize turns collected agent traces into smaller open-source models using the Tinker API, with optional closed-loop simulation training. wmo serve exposes an endpoint that routes requests between frontier and smaller models; on RouterBench, it maintains frontier quality at 27% lower cost. Rerun the pipeline as new traces arrive to continually improve a model you own.

🌐 Platform | 📚 Docs | Discord

Getting started

1. Register your providers.

pip install world-model-optimizer
wmo providers set

That verifies the provider and then offers to register its models as routing candidates in .wmo/pool.toml, the roster everything below chooses from. It searches the provider's own catalog (OpenRouter's 338 published models included) and asks only for what that backend needs. Re-run it to add another provider's models beside the ones already registered.

2. Tune a router on your OTel traces.

wmo build --file traces.jsonl --name my-model

# Score every registered model on held-out tasks from your traces
wmo optimize route sweep my-model --traces traces.otel.jsonl

# Deterministically reserve 30% for reporting and fit on the other 70%
wmo optimize route fit matrix.json --kind knn \
  --out .wmo/models/my-model/policy.json

3. Serve it.

wmo serve --name my-model

See what it bought you against the model you were using before. The report automatically excludes the router-fit scenarios recorded in the policy:

wmo optimize route report matrix.json .wmo/models/my-model/policy.json \
  --baseline gpt-5.5

Distill your own small model into the pool with wmo optimize distill, serve a single model with no routing via wmo optimize route pin, or build an optimized harness for your agent with wmo optimize harness.

Hosted platform

Create an account at platform.experientiallabs.ai, then authenticate the CLI:

wmo login

Copy an agent ID from the platform and run its current champion harness:

wmo run <agent-id>

E2B backend

Hosted agents already run in platform-managed E2B sandboxes. To evaluate a local optimization in E2B, install the extra and provide an E2B key:

pip install "world-model-optimizer[e2b]"
export E2B_API_KEY=...
wmo optimize harness my-agent my-environment --tasks tasks.jsonl --backend e2b

Use a world model as an API

world-model-optimizer includes world models that can be used to simulate your agent environment for testing and optimization.

from wmo import Action, ActionKind
from wmo.config.store import WorldModelStore
from wmo.engine.loader import load_world_model

model_dir = WorldModelStore(".wmo").resolve("airline")
wm, _provider = load_world_model(model_dir)

session = wm.new_session(task="check out the cart")
obs = wm.step(session.id, Action(kind=ActionKind.TOOL_CALL, name="add_to_cart",
                                 arguments={"sku": "A1"}))
print(obs.content)

Or over HTTP (same code path), namespaced by model name: GET /world_models, then POST /world_models/{name}/sessions and POST /world_models/{name}/sessions/{id}/step.

Run after platform login

After wmo login, the same wmo run command can open a hosted world model or run an agent's current champion harness in E2B. The platform manages model and sandbox credentials, so hosted runs do not need local API keys.

wmo login
wmo run <world-model-or-agent-id>
wmo run <agent-id> -u . --task "fix the failing tests"

Workspace upload is opt-in with -u: WMO live-syncs changes and preserves concurrent local edits. Long-running agents can detach, continue in the platform, and be messaged or reattached later.

wmo run <agent-id> -u . --detach
wmo run --send "Now run the full test suite"
wmo run --attach
wmo run --end

Runtime agents and optimizers in E2B sandboxes

WMO can run the real pi worker inside isolated E2B sandboxes while the world model supplies the environment. Optimization and evaluation rollouts run in parallel, and model credentials stay outside the sandbox.

wmo optimize harness my-agent my-environment --tasks tasks.jsonl --backend e2b
wmo eval tasks.jsonl --mode closed-loop --harness my-agent --harness-backend e2b

The optimizer can change prompts, tools, policies, skills, and runtime code. Every candidate is measured against the same simulated tasks, and only changes that pass the evaluation gates become the new versioned champion harness.

Development

Managed with uv; linting/formatting with ruff; type checking with ty; tasks run through just (brew install just / cargo install just). Conventions live in AGENTS.md.

just setup               # first time: .env from the template + uv sync
uv sync --extra dev      # env + dev tools
uv run ruff check .      # lint
uv run ruff format .     # format
uv run ty check          # type check
uv run pytest -q         # tests

Usage telemetry

wmo uses anonymous usage telemetry to track the volume of usage. Telemetry is strictly metadata. It never includes prompts, traces, actions, observations, file paths, model names, provider credentials, or raw user content.

Telemetry is enabled by default. To opt out for a project:

uv run wmo config telemetry disable

This writes .wmo/settings.toml. You can re-enable it with uv run wmo config telemetry enable, check the current setting with uv run wmo config telemetry status, or disable it for a process with DO_NOT_TRACK=1 or WMO_TELEMETRY=0.