Replays from the Agent World environment: a deterministic 2D office where agents fetch crates and deliver them to the goal zone. Same episodes, same frames, same rewards our eval harness logged — the viewer just plays them back.
| baseline | success% | mean steps | invalid rate |
|---|---|---|---|
| scripted | 100.0 | 25.10 | 0.0000 |
| prompted-llm-qwen0.5b | 0.0 | 120.00 | 0.4846 |
| prompted-llm-qwen7b | 100.0 | 45.90 | 0.0087 |
The 7B row's 100% team success is entirely its scripted teammate's work — across all 20 episodes the 7B agent delivered 0 of 60 crates. The 0.5B fails the same episodes by grabbing a crate and never releasing it.
Agents act through seven JSON tool calls — move, pick, place, inspect, msg, use, wait — one action per agent per step, with invalid actions becoming no-ops that surface an error in the agent's next observation. Rewards are exact geometric verifiers: +1 to the placing agent the first time a crate enters the goal-zone radius, with no learned or LLM judging anywhere. Episodes are deterministic — the only randomness is seeded spawn jitter, so the same seed and actions reproduce a byte-identical frame log, which is what this page replays. Baselines above are the committed 20-seed evaluation from the environment repo.