Tracked topic
… Across 14 MLLMs, even the strongest model reaches only 65.6% transition validity and 66.3% requirement coverage, and visual quality is no proxy for behavior Qwen3.6-35B-A3B on Markdown: V=80.8 yet T=15.5 . …
… Opencode, Claude Code, Copilot CLI, Gemini CLI, etc Agent Definition – a file Markdown/JSON that describes how an agent behaves; serves as a blueprint from which instances are created like a class in Java Agent Instance – a running process created from an Agent Definition like an object instantiate… …
… Opencode, Claude Code, Copilot CLI, Gemini CLI, etc Agent Definition – a file Markdown/JSON that describes how an agent behaves; serves as a blueprint from which instances are created like a class in Java Agent Instance – a running process created from an Agent Definition like an object instantiate… …
… Ships 65+ Markdown-defined skills, 5 end-to-end workflows, and a 3-stage evidence-to-claim audit cascade. …
… Further analyses show that the learned curator produces more targeted skill use, while the skills in SkillRepo evolve into more richly structured Markdown files that encode higher-level meta-skills over time. …
… An agent skill can be a Markdown instruction file, a Python script, a workflow definition, references, or a bundle that combines all of these. …
… On WideSearch we hit 38.50 Avg@4 Success Rate , 7.5× the previous SOTA and also generalise to XBench-DeepSearch at 73.0 accuracy. the thing that hooked me most here is the run verify reflect loop and the memory-driven bi-level setup, where a top-level orchestrator learns decomposition while a pool … …