Agent frameworks & orchestration
How enabling two settings tripled our scores on the ARC-AGI-3 benchmark
A benchmark score can reflect harness settings as well as model capability: OpenAI reports that retaining reasoning and enabling compaction raised GPT‑5.6 Sol’s public-set score from 13.3% to 38.3%.
Why it matters: When evaluating or deploying agents, API settings and harness design can materially affect measured performance. The article recommends comparing models with settings that match real-world use in ChatGPT and Codex.