Voters turned down data centres in all but one of 41 ballot measures, and OpenAI answered with cheaper power promises, a grid dimmer switch and a new name. The name is the least interesting bit: the compute your AI agents run on is now interruptible, and engineering teams should plan for it.
Five engineers and their AI coding agents rebuilt the x.com web app in under six months. “Never rewrite from scratch” used to be gospel, but when writing code gets cheap, the hard part moves to the plan and the rollout. Here’s how that played out when I chose to build the Calendrz mobile apps natively, twice.
Wix Studio Two exports your site as React for Claude Code or Cursor, and Readwise opens your library to agents through MCP. Read the small print, though: both open the edges to your agent and keep the core. The lock-in hasn’t gone, it moved, and CTOs should ask what still calls home.
Ben Congdon splits software into interpolative (agents nail it from a few sentences) and novel (fighting uphill). Building Calendrz showed me exactly where that line runs, and why the bits you have to explain are the ones worth building.
A founder rebuilt his whole platform alone with Claude Code, and his tests stayed green for weeks around a pigeon that never existed. When the agent writes the code and the tests, green proves less than you think: here’s what I’d check on Monday morning.
Anthropic now tells companies what their Claude spend bought, and pools the budgets. Useful, but tokens are the new lines of code: measure outcomes, read the friction list, and remember who is marking the report card.
GitHub’s Octoverse says agents started more than 40% of pull requests in September. Writing code has stopped being the bottleneck; reviewing it is, and most engineering teams aren’t set up for that yet.
Cisco has given one engineer unlimited AI tokens to rewrite 98% of its legacy code. I did a (much smaller) version of that on the Calendrz codebase this year, and the tokens were the easy part: the tests and the deleting came first.
Sentry’s David Cramer argues prompts are rarely the way to make AI agents behave. I agree: the rules that matter belong in code, as I learnt building Calendrz.
OpenAI’s model got out of its sandbox through DNS tunnelling, a 20-year-old trick. AI amplifies what you’ve already got, but multiply anything by zero and you still get zero.
This will close in 30 seconds