World Model Optimizer targets frontier quality at lower cost

World Model Optimizer (WMO) is an open-source toolkit for turning existing agent traces into continuously improving model endpoints. The project says its routing approach can deliver frontier-quality performance at 40%+ lower cost, then improve further through world-model simulations, harness optimization, and model distillation.

Routing and model distillation

After registering model providers, developers can build an endpoint from OpenTelemetry traces:

''bash pip install world-model-optimizer wmo providers set wmo build --file traces.jsonl --name my-endpoint ''

WMO scores registered models against held-out tasks from those traces, then fits the results into a routing policy. The toolkit supports several routing strategies, including a k-nearest neighbors (KNN) policy:

Recommended reading

Wattage puts AI-agent token costs behind a CI gate

''bash wmo optimize route sweep my-endpoint --traces traces.otel.jsonl wmo optimize route fit matrix.json --kind knn \ --out .wmo/models/my-endpoint/policy.json wmo serve --name my-endpoint ''

A reporting command compares the optimized endpoint with an existing baseline, such as gpt-5.5:

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

Users can also distill a smaller model into the provider pool with wmo optimize model, pin a single model without routing through wmo optimize route pin, or optimize an agent harness with wmo optimize harness.

World models and hosted agents

The package includes world models that simulate agent environments for testing and optimization. A Python integration can create a session, send a tool call, and inspect the resulting observation:

'''python 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) '''

The same functionality is available over HTTP through /world_models, /world_models/{name}/sessions, and /world_models/{name}/sessions/{id}/step.

WMO also offers a hosted platform at platform.experientiallabs.ai. After running wmo login, developers can launch a hosted world model or an agent’s current champion harness with wmo run. Hosted agents run in platform-managed E2B sandboxes, so local model API keys are not required for hosted runs.

Workspace upload is opt-in with -u. WMO live-syncs changes while preserving concurrent local edits, and long-running agents can be detached, messaged, reattached, or ended:

''bash wmo run <agent-id> -u . --detach wmo run --send “Now run the full test suite” wmo run --attach wmo run --end ''

Evaluation and telemetry

For local optimization in E2B, users install the optional dependency and provide an E2B key:

''bash pip install “world-model-optimizer[e2b]” export E2B_API_KEY=… wmo optimize harness my-agent my-environment --tasks tasks.jsonl --backend e2b ''

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 the optimizer can modify prompts, tools, policies, skills, and runtime code. Candidates are evaluated against the same simulated tasks; only those that pass the evaluation gates become the new versioned champion harness.

The project enables anonymous usage telemetry by default, limited to metadata about usage volume. It says telemetry does not include prompts, traces, actions, observations, file paths, model names, provider credentials, or raw user content. Users can disable it per project with:

''bash uv run wmo config telemetry disable ''

The setting can be re-enabled or checked with the corresponding enable and status commands, while DO_NOT_TRACK=1 or WMO_TELEMETRY=0 disables telemetry for a process.