AI digest: agents everywhere
Everyone's building AI agents now, from CopilotKit's new stack to Qwen's million-token model.
The agent wave is properly here. This week brought production tools, massive context windows, and the first AI proof to land in a top maths journal.
CopilotKit ships production agent infrastructure
CopilotKit released their 2026 stack with AG-UI protocol, AIMock testing suite, and Pathfinder server for agentic AI development. This matters because most agent frameworks are still research toys. Having proper testing and deployment tools could finally make agents workable in production.
Qwen3.7-Max brings million-token reasoning
Alibaba’s new Qwen3.7-Max packs a 1M token context window with extended thinking modes for long-horizon tasks. Scored 56.6 on the AI Intelligence Index, ranking fifth among proprietary models. The context size is impressive but we’re still waiting to see if these massive windows actually work reliably in practice.
Cohere goes big with Command A+
Cohere open-sourced Command A+, a 218B sparse MoE model that runs on just two H100s with quantisation. It’s their first multimodal reasoning model and supports 48 languages. The hardware efficiency is genuinely impressive, though sparse MoE models can be unpredictable when you push them.
OpenAI solves 80-year-old maths problem
An OpenAI reasoning model disproved an Erdős conjecture from 1946 using algebraic number theory techniques that surprised human mathematicians. This is the first AI proof worthy of a top-tier maths journal. Shows these models can genuinely discover new mathematical insights, not just rehash existing knowledge.
ByteDance unifies everything in 3B parameters
ByteDance’s Lance model handles image and video understanding, generation, and editing in a single 3B parameter framework. The efficiency is remarkable but we’ll see if unified models actually outperform specialised ones in real applications.