Sandboxed harness for testing coding-agent learning
Arc Code tests whether a general coding agent can learn unfamiliar ARC games by creating its own parsers, simulators, and search programs during play.
From Github Awesome — GitHub Trending Weekly #45: comet, blobatar, cumora, herdr, barehands, ip-as-logo-skill, Aura, md2hd at 06:27
Problem: Agent evaluations need to test tool creation and adaptation rather than only predefined task execution.
For: Researchers evaluating general coding-agent capabilities.
Examples
- ARC games: the tasks used to test whether a general coding agent can learn unfamiliar environments.
Soon you can unlock the full business plan.
Behind this: 5 build steps · how to validate demand · 1 thing the video never answers.