World Model Optimizer Review 2026: Turn Agent Traces Into Cheaper Models You Own

Most teams throw away their agent traces. World Model Optimizer’s argument is that those traces are the most valuable training asset you have — a record of exactly the tasks your agents actually face, with frontier-model answers attached.

wmo turns that record into a smaller model you own, plus a router that decides when the frontier model is worth paying for.

What is World Model Optimizer?

World Model Optimizer (wmo) is a Python CLI from Experiential Labs. It does two connected things:

1. wmo optimize distills collected agent traces into smaller open-source models via the Tinker API, with optional closed-loop simulation training.
2. wmo serve exposes an endpoint that routes each request between frontier and smaller models. On RouterBench, the project reports it maintains frontier quality at 27% lower cost.

The loop is the point: rerun the pipeline as new traces arrive and the model you own keeps improving.

How the workflow looks

pip install world-model-optimizer
wmo providers set                       # register providers and routing candidates
wmo build --file traces.jsonl --name my-model
wmo optimize route sweep my-model --traces traces.otel.jsonl
wmo optimize route fit matrix.json --kind knn --out .wmo/models/my-model/policy.json
wmo serve --name my-model

One detail worth calling out for anyone who cares about honest evaluation: the fitting step deterministically reserves 30% of scenarios for reporting and fits on the other 70%, and wmo optimize route report automatically excludes the router-fit scenarios from the comparison. That’s the right default — it makes the reported savings harder to fool yourself with.

Key features

• Distillation from your own traces — build a small model that reflects your actual task distribution, not a generic benchmark.
• Router fitting on OTel traces — score every registered model on held-out tasks, then fit a routing policy.
• Broad provider registry — searches each provider’s catalog, including OpenRouter’s 338 published models.
• World models as an API — simulate your agent environment in Python to test and optimize before touching production.
• Hosted platform — wmo login, then run a managed agent’s current champion harness, with E2B sandboxes for evaluation.

Who should use it?

This is a tool for teams with volume and traces. If you’re running agents at enough scale that inference spend is a line item, and you have OTel traces to learn from, the economics are compelling — a distilled model you own has no per-token markup.

It’s a poor fit early on. With a handful of traces, distillation has nothing to learn from and routing has nothing to route. Get to real traffic first.

If you want cost reduction without a training pipeline, a plain router like LiteLLM or Millwright gets you part of the way with far less work. wmo goes further precisely because it trains a model, which is also why it costs more effort. For a broader look at self-directed research tooling, Open Science sits in adjacent territory.

Pros and cons

Pros: turns a wasted asset (traces) into an owned model; honest held-out reporting built into the workflow; routing and distillation in one tool; simulation environment for closed-loop testing; broad provider support; hosted option if you don’t want to run it.

Cons: no license file declared on the repository, which matters for commercial use; depends on the Tinker API; needs meaningful trace volume before it works; the 27% figure is the vendor’s own RouterBench result; young project with a high open-issue count.

Pricing

The CLI installs free from PyPI. A hosted platform is available at Experiential Labs for managed agents and E2B-backed evaluation. Note that the repository does not declare a license — worth clarifying before you build a commercial workflow on it.

FAQ

What data do I need to start? Agent traces, ideally in OTel format. The quality and volume of those traces set the ceiling on what distillation can achieve.

Does it replace my frontier model? No — it routes. Hard requests still go to the frontier model; easy ones go to the small model you distilled. That’s where the savings come from.

Is the 27% figure trustworthy? It’s the project’s own RouterBench result. The methodology (30% held out, fit scenarios excluded from reporting) is sound, but you should reproduce it on your own traffic before budgeting against it.

Can I use it commercially? Check first. The repository has no declared license as of this review, so the terms are ambiguous.