Labrats — Tiny lab agents in DiscoveryWorld

A small (≤32B) LLM agent inside the DiscoveryWorld simulator, with a two-tier memory layer (private episodic + shared lab notebook).

Run a live episode — spin up N ReAct agents on a fresh DiscoveryWorld scenario. Choose the HF backend (Hugging Face Inference Providers, needs an HF_TOKEN) or the local backend (in-process llama-cpp, needs the LLAMA_* env vars). The run streams into a new trace that appears in the Episode replay tab when it finishes.

Status: HF_TOKEN detected, no local model.

Backend

local hf

Scenario

Difficulty

Seed

Agents

1 4

Max steps

1 120

Memory

Dialogue

Model (HF backend)

Max tokens / step

Reasoning models need room to think + emit the action.

256 4096

Idle — configure a run and press Start live episode.

(frames appear here once a run starts)

Out-of-band peer-to-peer messages between agents, newest at the bottom. The latest step is highlighted in dark blue.

📓 Lab notebook (shared)

(empty)

🗒️ Private notes

(empty)