PRO-LONG: Keep the Full Log and Let a Coding Agent Grep It — +18pp on ARC-AGI-3 With 4.2–5.8× Fewer Tokens

arXiv 2607.20064·high signal

Instead of summarizing or evicting history, PRO-LONG keeps a complete structured interaction log as 'programmatic memory' and uses coding-agent search to query it, sidestepping the usual context-management tradeoff where preserving more information makes retrieval harder. On the full ARC-AGI-3 public game set it beats a base coding agent by 18.0 percentage points across frontier models and matches or exceeds specialized harnesses (up to 76.1% pass@1) while spending 4.2–5.8× fewer tokens; with Fable 5 it reaches 97.4% best@2 for $1,750 total. Code is public at github.com/alexisfox7/PRO-LONG — a directly copyable pattern for long-horizon agent harnesses.