Ultrafast Agents, Mistral Large 4 & OpenAI’s Math Release | WeeklyClaw #33

What this episode covers

• Mistral Large 4 previews 1T total / 52B active parameters, with weights promised for the end of October.
• Reflection Beam previews 501B total / 23B active; weights, technical report and model card promised later in the month.
• OpenAI releases 722 math manuscripts with 300 of 719 top-line results formalized; the separate Alman–Williams paper credits Claude with a faster algorithm.
• Microsoft brings hybrid agent execution to Windows 11: Execution Containers reach GA with OpenClaw support.
• ThinkingBox finds two-thirds of 121,680 agent trials failed to leave the right state behind.
• Google announces a shared Gemini workplace agent with persistent cloud execution and coworker identities.
• Nous Research announces a $90M raise and a plan for Hermes for Businesses.
• Anthropic’s usage-policy update prohibits sustained cruel behavior toward models; LMCache discloses an unauthenticated RCE.
• The closing debate asks whether polished consumer UX pulls users away from open, model-agnostic agents.

Published record