AI Field Notes: Google, Lets, and more

The IDE maker just dropped an SDK update that can pull your agent back on track. If an agent heads off track, “run.steer()” lets you slip in notes mid-turn to course-correct your agent. Busy subagents seamlessly slip into the background to keep working, then report straight back to the main agent. You can also tag cust

A 19-year-old wants to give you AI intelligence you can actually own. His a16z-backed hardware startup, Ghost, just unveiled Core, a screenless box that runs models locally. Ghost says your data stays encrypted so even the company can't read it. You can also swap in Hugging Face models or plug in harnesses like OpenCod

Also: 4 tests every engineer should run on AI-generated code, fun projects for you to build, and the pioneer of open-source claims only “one kind of software” will survive the age of AI.

A product designer at Warp built a physical answering machine for his family with Astra. The result has devs on the internet obsessed (1.6K bookmarks).

Open-source pioneer Eric Raymond claims “software secrecy” is nearly dead, and only one kind of software can still keep its secrets (7.4K likes).

Karpathy says custom AI explainer videos will change education. One dev put it to the test with a video explaining how Shazam (the music app) works, and the results are stunning (8.1K likes).

This guide walks through the setup that makes parallel Claude agents actually useful: choosing the right context, models and effort levels, writing project instructions; configuring cloud environments and connectors; and keeping memory clean. The idea is to do the boring setup once so new threads can start working with

Guillermo Rauch breaks down how he’s building around cloud agents, faster feedback loops, and software factories where humans spend less time typing code and more time steering the work. The big shift is from writing everything yourself to orchestrating agents with better taste, judgment, and clear goals.