PRO-LONG hits 97.4% on public ARC-AGI-3 for $1,750 with a log.txt memory harness
GregKamradt · x · 2026-07-28
PRO-LONG reaches 97.4% on public ARC-AGI-3 for $1,750
The paper introduces PRO-LONG, a minimal memory framework for long-horizon LLM agents built around a structured log.txt and simple search over the full interaction history.
- On the public ARC-AGI-3 set, the team reports 97.4% best@2 with Fable 5 at a total cost of $1,750.
- Compared with prior harnesses, it uses 4.2–5.8× fewer tokens while matching or exceeding specialized baselines.
- The authors argue that long tasks benefit from preserving complete structured logs rather than squeezing everything into limited context.
- Code, logs, and scorecards are released alongside the paper.
Related event: PRO-LONG Framework Scores 97.4% on ARC-AGI-3(2 posts)→
More from coding & agent
- Claude Opus 5 keeps running Python scripts to edit files in coding workflows — samuelcolvin · 2026-07-28
- Vercel highlights its open-source stack, including Eve and agent-browser — evilrabbit_ · 2026-07-28
- LightOn says its agent search hits 86.27% accuracy with just 9.7 calls — IgorCarron · 2026-07-28
- The Modern Wait Equation: Fix AI Slop Now or Wait for the Next Model? — ilanbigio · 2026-07-28
- Kimi K3 pairs a 2.8T MoE with 1M context and multi-teacher post-training — novasarc01 · 2026-07-28
- Should an agent approval survive a resource change before execution? — marcelk231 · 2026-07-28