World Model Optimizer: Open-Source Tool Turns Agent Traces Into Cheaper, Smarter Models

A new open-source project called World Model Optimizer (WMO) aims to close the loop between agent instrumentation and continuous model improvement. By ingesting the OpenTelemetry traces you already collect, WMO lets you train a router that selects the cheapest capable model for each call, distill a smaller specialized model, or simulate entire agent environments with world models — all without manual tuning.

WMO is built around a CLI and hosted platform at experientiallabs.ai. The core workflow is straightforward: register your model providers (OpenAI, Anthropic, open-source endpoints, etc.), point WMO at your existing OTel traces, and it tunes a router that typically delivers frontier-level quality at 40% or more cost reduction. You can then serve that router as a drop-in replacement for your current model endpoint.

For teams that want to go further, WMO supports multiple optimization strategies:

• wmo optimize model — distill a smaller model and add it to your routing pool
• wmo optimize route pin — force a single model with no routing (useful for testing or when you need tight control)
• wmo optimize harness — build an optimized harness for your entire agent, modifying prompts, tools, policies, and even runtime code

A key differentiator is world model simulation. WMO includes world models that can simulate your agent's environment, allowing you to test and optimize changes without touching production. The optimizer evaluates each candidate against the same simulated tasks, and only changes that pass the evaluation gates become the new versioned champion harness.

For hosted runs, WMO uses E2B sandboxes to isolate the real agent worker while the world model supplies the environment. Model credentials stay outside the sandbox, and the platform manages all API keys — so hosted runs don't need local secrets. The CLI can also detach long-running agents, let them continue in the platform, and reattach later.

The project is managed with uv, uses ruff for formatting, and encourages contributions via a standard PR workflow. Anonymous usage telemetry is enabled by default but can be opted out via config or environment variables.

WMO is available now on GitHub under the experientiallabs organization. The platform is free-tier accessible, and the core optimizer is MIT-licensed.