FrontierCode 2.0 Tops Out at 45.5%. More Reasoning Effort Made Agents Change Code Nobody Asked For
Cognition released FrontierCode 2.0 on October 27. It's a coding benchmark that asks whether an agent's pull request would actually get merged by the people who maintain the repo. The best result,
Cognition released FrontierCode 2.0 on October 27. It's a coding benchmark that asks whether an agent's pull request would actually get merged by the people who maintain the repo. The best result, Claude Opus 5.5 running in Claude Code at max effort, scores 45.5%. The finding I keep coming back to is a different one. On these tasks, turning reasoning effort up often made agents change more code than anyone asked for. If you hand coding agents real tickets, this is a better map of where they go w
more from @Joshua_WD
YC Labs Ran AI Agent Swarms in Minecraft and Mindustry. What Multi-Agent Builders Should Copy
YC Labs put 50 AI agents in a Minecraft survival world for three hours, told them to build a civilization, and watched a boss emerge by accident. It then ran Claude Opus 5.5 and GPT-6.1 Sol teams
90% of US Work Hours Go to Tasks AI Benchmarks Barely Test, Mercor Says. Build Your Own Evals
Mercor mapped 29 public AI benchmarks onto the Labor Department's catalog of American jobs and found that about 90% of US working hours go to tasks with little or no benchmark coverage. Only 208 of