MF Goalkeeper
The Harness Engineering & Model Wrangling HackathonLong-running agents degrade because their context is a pile and they drift because the thing doing the work is also the thing remembering why. goalkeeper keeps everything the swarm learns in a raw, append-only library on MongoDB Atlas, keeps the goal small and written once by a human, and gives each stateless worker a context built fresh for its task from the library, about 4k tokens however large the library grows. A deterministic gate checks every result and a hidden score the workers never see reopens puzzles with a hint. Today it ran 400 ARC-AGI-1 puzzles on DeepSeek V4 Flash with thinking off. The same model one shot with no harness solved 10 percent. With the harness it solved 80 percent of the puzzles it finished, 74 by 16:30, for under a dollar, while the library grew to 18.7M tokens and every worker still read 4k. Live at goalkeeper-gamma.vercel.app
Team
Placement
Participant
