AI digest: Agents get serious infrastructure
Major breakthroughs in AI agent infrastructure, Google's automated research writing, and Z.AI's open-weight coding model.
This week brought proper infrastructure for AI agents and some genuinely impressive open-weight models. The research automation tools are getting scary good.
OSGym makes agent training actually affordable
Training computer-use agents just got 100x cheaper. OSGym manages over 1,000 OS replicas for just $0.23 per day, solving the massive infrastructure headache that’s been holding back agent research. This is the kind of boring but crucial work that actually moves the field forward.
Google’s PaperOrchestra writes research papers end-to-end
Google’s PaperOrchestra uses multiple AI agents to turn messy lab notes and experimental results into polished academic papers. The multi-agent approach handles everything from structuring arguments to formatting citations. We’re genuinely curious how long before this becomes standard practice in academic labs.
Z.AI drops massive open-weight coding model
GLM-5.1 is a 754B parameter model built specifically for agentic tasks, not just chat completions. It hits state-of-the-art on SWE-Bench Pro and can run autonomously for 8 hours straight. The fact it’s open-weight makes this particularly significant for anyone building serious coding agents.
Astropad builds remote desktop for AI agents
Astropad’s Workbench lets you monitor and control AI agents running on Mac Minis from your iPhone. Low-latency streaming designed for watching agents work, not fixing your mum’s printer. Smart pivot from their traditional creative tools business into the agent monitoring space.