LLMs as 5x Faster Sandboxes A new open-source tool called World Model Harness (wmh) uses frontier LLMs to simulate agent environments from OpenTelemetry traces, replacing traditional sandboxes. The tool ingests recorded state-action-observation steps, builds a retrieval index, and serves a world model that agents can interact with via a local HTTP backend. This approach claims to be up to 5x faster than standing up a full sandbox environment. Docker as an LLM.Simulate an agent environment from traces instead of standing up a sandbox. A frontier LLM acts as the environment your agent steps against, reconstructed from OpenTelemetry traces. The harness ingests recorded state, action - observation steps, builds a retrieval index, evolves the base environment prompt with GEPA, and serves the resulting world model locally. Build from OTel traces: ingest, normalize, split train/held-out, index the replay buffer, and optimize the environment prompt. Serve or play the built model: agents call WorldModel.step action in-process or through the local HTTP backend. Evaluate reconstruction fidelity with wmh eval against trace files. uv sync wmh providers verify wmh build --name airline --file examples/tau-bench/traces.otel.jsonl wmh list wmh eval examples/tau-bench/traces.otel.jsonl wmh eval list wmh eval run tau-bench wmh eval results wmh examples list wmh examples run tau-bench -- --trace 0 wmh serve wmh demo --name airline wmh play --name airline wmh build with no flags launches a guided creation wizard on an interactive terminal. Pass --file and related flags, or --no-interactive , for scriptable runs. World models are named and stored under .wmh/models/