Self-Improving Harnesses Push ARC-AGI to 95.5%

Prime Agent hit 95.5% on ARC-AGI using self-improving harnesses on Claude Opus, up from 30% on the private set without scaffolding.

The demo at YC Paper Club showed three live systems, Prime Agent, Open Jarvis and YC's QM, all rewriting their own prompts and code. They run as persistent local processes with file, REPL and sub-agent access, so the real risk is local session access, not remote exploit.

The full discussion can be found on YC's YouTube channel.

How self-improving harnesses actually work

The model on its own is a Turing machine, tokens in, tokens out. The harness wraps it in a von Neumann layer: L1 weights, L2 active context and REPL, L3 files and sub-agents, all of which it can garbage-collect and rewrite. DSPy evolves the system prompt through genetic search over small training sets. Darwin Godel Machines evolve the harness code itself and archive winners to breed stronger scaffolds.

Why this matters and what it still doesn't fix

For builders, the payoff is immediate. An 18% framework gain can decide whether Arcade AI ships, and Open Jarvis claims 800x lower cost by keeping inference on an Mac Mini-class device. Qwen 3.8 27B now matches Claude 4.6 Opus from August 2025, shrinking the cloud lead that local harnesses rely on. Nothing here is patched by default. QM centralizes chats in Postgres and gates writes with human review, but auto-approval tends to drift and weak permissions leak secrets, so isolate stores and audit every rewrite.

Self-editing scaffolds work. They just only work inside the boundary you define.