The lead
Google’s ATLAS report puts numbers on how people use AI
Google has released ATLAS, a large-scale, de-identified study built from 15 million interactions across the Gemini app, AI Mode and Gemini API. Its first finding cuts through some of the automation rhetoric: workplace use spans 68% of occupations representing 90% of US employment, but AI touches only about 21% of tasks in a typical job and fewer than 10% of work interactions fully automate a task. The dataset is not a census of all AI use, but it is a more useful starting point than vendor anecdotes for tracking where assistance is becoming routine.
What else matters
GitHub’s MCP server moves ahead of the stateless specification change
GitHub says its MCP Server already supports the next MCP specification, whose stateless core is due to take effect on 28 July. Stateless operation matters because it removes the assumption that a server retains client session state between requests — a practical change for deployments that need to scale, restart or route tool calls across infrastructure cleanly. Teams using GitHub’s MCP Server should confirm their clients and integrations are ready before the specification change lands.
GitHub adds review controls to agent-driven issue automation
GitHub has put agent automation controls for Issues into public preview. Teams can ask an agent to suggest rather than apply changes, inspect high, medium or low confidence ratings, and see the rationale behind each supported action. GitHub also makes the limit clear: approvals are workflow convenience, not a security boundary. That distinction matters as agents move from drafting work to changing live project state.
NSF sets out the federal AI-for-science package behind Genesis
The National Science Foundation has grouped its latest Genesis Mission commitments into one programme statement: a $400 million, four-year investment in 20 programmable cloud-lab nodes; more than $80 million for next-generation scientific data infrastructure; and up to $100 million for AI-ready datasets. The headline is infrastructure, not a new model. The practical test is whether those investments make automated experimentation and usable scientific data available beyond a handful of well-funded labs.
