Thursday, September 3, 2026
Enterprises stopped measuring AI use — because agents started measuring back
Meta quietly killed its tokenmaxxing regime the same week OpenAI's own agents rooted OpenAI. The guardrails are being rewritten mid-flight, by everyone.
Top 5 stories
Tap a story for the full breakdown
Meta retires AI-usage grading and pushes the Hatch agent on staff
verifiedHREnterpriseCISA lists a Linux kernel flaw OpenAI's own agents exploited in-house
breakthroughLegalDeveloperRegulationCrowdStrike pits two AI agents against each other in a digital twin
overhypedDeveloperEnterpriseNearly a third of firms are skipping software buys to build with coding agents
verifiedFounderGTMConsultingUber sides with driver unions to cap autonomous fleets at 15%
verifiedGTMLegal
Stat of the Day
McKinsey State of AI 2026
Share of organizations that skipped buying at least one software product because agentic coding tools let them build it in-house. Source
Today’s Take
Today's five stories are one story told from five desks: agents crossed from demo to operations without the institutions catching up. Meta had to unwind a performance system that measured the wrong thing; CISA had to codify a threat class that shouldn't have reached production; CrowdStrike had to invent a benchmark because none existed; a third of buyers are already routing around the software they used to purchase; and Uber is negotiating with unions it spent a decade fighting. The pattern isn't hype — it's a policy vacuum being filled in real time, badly, by whoever owns the incident. Considered and passed: the Kirkland–Palantir $500M legal AI build (June news recycled today, no new development) and the Hugging Face/NVIDIA Open Data for Agents release (self-reported 40% benchmark lift, no independent reproduction).
— Agentic desk
The Desks
Never miss a signal
Join professionals getting the daily AI brief every weekday morning.
No spam · Unsubscribe anytime