An AI agent got into an Australian government system in June. The government found out in September, from an email to an inbox someone checked once a day.
Read →In January, Tailwind laid off three-quarters of its engineers while the framework grew faster than ever. Traffic to its docs was down about 40%, revenue close to 80% — because the docs were the only way customers found the paid product. In September the company joined Shopify, and OpenAI began selling the visit back as an ad. If your website finds you customers, count how much of your revenue depends on someone arriving.
In July, Robert C. Martin — the man who taught a generation to write readable code — said he no longer reads what his agents write. He replaced reading with a gauntlet of tests and metrics. Grady Booch had already objected that metrics confirm function, not safety. Z.ai, reporting on its own model, then found flaws averaging 26.6 years old, the oldest from 1981. Readable code was never the product. Confidence was — and you now buy exactly as much as your gauntlet can measure.
Since 2 August, EU law requires AI-generated text to be machine-readable as such, and Anthropic complied worldwide rather than only in Europe. The mark Anthropic ships does not say a machine wrote the text. It says a machine touched it — proofreading and translation trigger it too, and a detection is by Anthropic's own account not fully conclusive. Procurement is already writing AI-disclosure clauses. Decide now what your business lets a model touch.
On 14 August, auto mode becomes the default in Claude Code. Anthropic's study found human review caught 13.6% of dangerous commands against the classifier's 89%; separately, users approve 97% of permission prompts. The click was never the control. What you authorise before the session starts is — and that decision belongs to an owner, not an engineer.
Ford put 900 AI cameras on its quality line, then hired back 350 veteran engineers to retrain them. A third of the US hiring managers who cut a role for AI have already rehired for it. The pattern is always the same: automation takes the routine work, and the routine work was where your experts came from.
Half the web goes dark on the same afternoon because it all rents from the same three companies — and each has proven it can take itself offline with a routine change nobody attacked. You can't shop your way to a safer landlord. You can decide, in advance, what your business does on their bad day.
In June, Stripe ran a codebase-wide migration through a fifty-million-line Ruby codebase in a single day. Execution just stopped being the expensive part of changing software. What remains — the part almost nobody measures — is how much change your verification machinery can absorb.
AI tools change every few months; the operator's habits don't. The two that decide whether a session was worth your time are getting the right context in and verifying what comes out. Both are skills the model can't do for you.
A small website can sit on the internet for $0 a month, load faster than the paid version, and survive a traffic spike without going down. Here's what changed, what you give up, and why most people still pay $30 a month for the slower option.
Vite+ rolls runtime, package manager, dev server, build, test, lint, format, and monorepo orchestration into a single binary. The interesting question is not whether the consolidation works — it's what you give up by taking the bet.