Two ex-Datadog engineers just raised $7M to build a router that picks between Claude Code, Codex, and open-source models per project. The pitch: don't bet your whole dev stack on one lab. A company now exists whose entire product is selling you an exit. Makes sense!
Every few months a model ships that writes code better than the last one.
None of them ship better at deciding what to build.
That skill has compounded for 30 years and will compound for 30 more. Hire for it. Become it.
Apple's line yesterday: "privacy in AI is non-negotiable." The part worth noticing wasn't the promise. It was "outside experts can verify this at any time." Everyone in AI says trust me. Almost nobody says check me.
Vendor risk used to mean "what if they raise the price." In 2026 it means "what if they delete the product." Gemini CLI got about a month's notice before it stops serving requests. Build your stack like every tool in it has a sunset date, because at least one of them does.
Grok Build shipped with the Agent Client Protocol. Quietly the best part of the launch. It means you can swap coding agents like you swap editors. The winners of the agent era won't be locked in. They'll be the ones who can leave.
A paper this week is literally titled "The End of Software Engineering." Read the actual claim and it's narrower: code became disposable. Fine. Code was never the engineering. The engineering was deciding what the code should do.
The real trick in OpenAI's new ChatGPT memory isn't remembering more. It's pruning what went stale. Which is exactly the skill your team's wiki has never had. Every doc you wrote and never deleted is a memory leak with good intentions.
Microsoft spent seven years buying its AI brain from OpenAI. This week it shipped its own and is swapping it into Copilot for millions of devs in August. The biggest build-vs-buy call in software just got made in public. Most teams never even chose their model.
GitHub Copilot went from a $29 flat fee to metered billing this week. Some devs report bills jumping past $750. The price didn't change. The ceiling did. We were never paying for the tool. We were paying for the illusion that running it was free.
Anthropic is holding 10,000 findings under a 90-day disclosure clock. Sit with the pressure that puts on maintainers. We built AI that finds flaws at machine speed and handed the fixes to humans on calendar time. The mismatch is the whole story.
Yesterday's Codex tax case study is described as "self-improving." Read the architecture. The "self" is a team of practitioners writing corrections into a harness. The "improving" is engineers maintaining the eval suite. The model is a participant, not the protagonist.