Today's Overview
Six major vendors launched enforcement layers for AI agents this week-Devenex, Antigravity, Notion, Claude, OpenAI, and Salesforce. Each shipped a different execution control plane. None shipped the policy file your agent repo needs to use them.
The Policy File Gap
An analysis of this week's launches identifies the problem clearly: enforcement infrastructure exists now. What doesn't exist is a standard way to define what agents are actually allowed to do. Policy files need four sections: action classes (read, write, send-external, transact, escalate), blast radius caps (write limits, transaction ceilings), escalation triggers (who gets notified when caps are hit), and evidence schemas (what gets logged for audit). This isn't vendor-specific; it's the shape every control plane will enforce against. And it lives in your repo, not theirs. Teams deploying agents this month are writing this themselves, from scratch, for each vendor's system. By next month, that's going to look like negligence.
On the iOS Side, On-Device Gets Real
Apple's Foundation Models framework launched at WWDC with something developers have wanted for years: language models running entirely on-device, no API calls, no cloud dependency. A 3-billion-parameter model now fits in Swift with a @Generable macro that turns your Swift structs into type-safe AI outputs. No per-token costs. Complete privacy. The catch: A17 Pro+ hardware only. But for builders targeting recent iPhones, this changes the maths on every local-inference feature you've been hesitant to ship. The template is already in the wild-teams are shipping proof-of-concepts by end of week.
Google also made noise on the design side this week, positioning itself as a contender in AI-powered design tools at I/O 2026. And YouTube's search just turned conversational, with Gemini Omni integration letting you ask video questions instead of typing keywords. Meanwhile, on the cost side, Google's releasing Gemini 3.5 Flash to help enterprises spend fewer tokens per request-a direct response to token-cost fatigue.
Quantum Keeps Going ahead
Three papers on arXiv this week show quantum computing moving from toy problems to practical infrastructure concerns. Statistical quantum phase estimation now handles negative Pauli weights and changepoint detection without requiring prior knowledge of ground-state overlap. State preparation compilation shows sampling-based methods retaining advantages even after accounting for total gate overhead. And asymptotic-preserving methods for open systems prove that layered quantum protocols can solve multiscale physical systems with explicit error bounds. None of this makes quantum computers suddenly useful for real work tomorrow. But all of it removes excuses for not building the infrastructure that will handle them when they are.
The pattern across the week: enforcement, privacy, and cost efficiency. Agents need policies before they scale. iOS developers can stop asking permission to run AI locally. Google's betting design is the next frontier. And quantum researchers are solving the engineering problems that come after the physics works. None of this solves the hardest problem-what should agents actually do-but they've all stopped pretending it's not a problem to solve.