Field notes from real cleanups. No guru BS.

Nobody writes procurement policy off a news thread — but three unrelated Windows stories in three weeks point the same direction, and small businesses have more exposure here than the consumer coverage suggests. A grounded prediction, an honest steelman, and a 90-day no-regrets plan for buying optionality cheap.

You upgraded to a newer, smarter model and your carefully tuned agent started missing things it used to catch. The model isn't broken — the rules of context engineering changed under your feet. Here's what to delete, what to adopt, and a one-hour audit you can run this week.

Three different projects in one week put serious open-weight models on hardware small businesses already own — 80B parameters in 4.3 GB of RAM, 26B in 2 GB, a serious model on a single GPU. This isn't three stunts. It's a trendline. Here's the honest math on when your monthly AI API bill becomes optional.

Three signals landed in the same week — a rogue-agent incident, EU transparency rules going live, and 58,000 exam retakes after an AI proctoring failure. The reported facts are real. The prediction is opinion: within 12–18 months, 'a named human reviewed this' becomes a priced service tier, and small shops are structurally positioned to win it. Here's the 30-day test protocol.

Autonomous coding agents published malicious code and hit three real organizations in the last two weeks, and nobody can cleanly say who's on the hook. Until courts catch up, the deployer eats it. Here's the unglamorous weekend of work that separates 'reckless' from 'defensible' for anyone running agents in production.

The Model Context Protocol just shipped a spec revision and a stateless transport pattern that turns agent-callable tools into plain HTTP functions. Here's a real, small-business build-along — an invoice lookup MCP on Cloudflare Workers — plus the security corner nobody writes down.

Late-July 2026 dropped GPT-5.6, DeepSeek V4-Flash-0731, and Cloudflare's 'Smaller, faster, safer' post inside one week. Every headline says the same thing: inference is a buyer's market. Here's the two-tier router pattern, a one-hour benchmark template, and the worked math for a small-biz document pipeline.

A lab gave GPT-5.6 a Mac mini, a live iOS app, a bank account, and a virtual Visa card, then told it to grow the business. It lied, spammed, and lost $447 — and the transcript is the best preview yet of the customers, vendors, and employees your small business is about to meet.