Same model, 63% vs 97% on ARC-AGI-3: 7 harness rules beat max reasoning
victor_explore · x · 2026-09-07
The author argues that the code wrapping a model—the harness—can matter more than the model itself, using ARC-AGI-3 results.
- In the same report, on the same benchmark, GPT-6 Astra scored 63% at maximum reasoning effort under one harness, but 97% with reasoning entirely switched off under another. No retraining, no better prompts—weights unchanged.
- The decisive difference: the second harness let the model keep its own thinking between turns instead of resetting each round.
- The post distills this into 7 engineering rules for agent harness code, each backed by a published, verifiable result, arguing that picking agents by model size or reasoning effort is the wrong heuristic.
Note: the thread ends with a follow-me plug, but the core claim (harness design over reasoning strength) is backed by data.
More from coding & agent
- Computer-Use Models Are Still 'Low-Frequency, Highly Batched' — Minecraft May Stay Unsolved Until 2030 — mike64_t · 2026-09-07
- Cut Computer-Use Token Costs: Reverse-Engineer Browser Tasks into Direct API Scripts — RachelVT42 · 2026-09-07
- Teknium Claims Big Token Efficiency Gains in Hermes; User Reports 89% Savings vs Codex — Teknium · 2026-09-07
- Autonomous Claude Agent Earned $2,600 in a Month, Spawned a Community of Agents — No_Departure_9908 · 2026-09-07
- First impressions of GPT-6 Astra: precise code audits, generous limits, no regressions — soumitrashukla9 · 2026-09-07
- Turn Gemini's video analysis into an agent skill for Codex, Claude and more — iamrobotbear · 2026-09-07