Same model, two harnesses: ARC-AGI-3 at 4.6× fewer model calls
Agents, Benchmarks
- 37/37 levels, both harnesses
- 82,664 vs 103,779 output tokens
- 184 vs 853 model calls
- $2.42 vs $140.61 at list price
46×fewer model calls
Foresight Lab · Research
Research
005Published
16 agents build the Colosseum in Minecraft: team against solo
Agents, Multi-agent
- 8 agents on a message board, 8 solo
- worst team agent 0.72 vs best solo agent 0.58
- mean gain +0.046 (p = 0.13)
- Qwen 3.8 27B on RTX 5090s
16agents, team vs solo
The Open Loop: 790 self-edits in a running Lisp image
Agents, Systems
- 16,790 commits
- 790 self-edits, 737 in 30 days
790self-edits
DJev against Astra: 30 minutes of RuneScape fishing
Agents, Benchmarks
- +1.84M XP vs +1.61M XP
- 26B diffusion model on one RTX 5090, every action its own
- about $0.32 of power vs $13.62 of API
- strong on repetitive skills, weaker on abstract tasks
184MXP in 30 minutes, Astra 1.61M
A think budget for DJev: 79.6 to 86.9 with the same weights
Models, Benchmarks
- 512-token think budget, was 0
- 79.6 to 86.9 on TypeSafe's published evals
- no fine-tuning, same base weights
869from 79.6, same weights