Just a few links this week.

How the Claude Code team does code review

A useful follow-up to my last post on code ownership. The Claude Code team describes how it handles the volume of agent-written code with automated reviews, checks matched to risk, and fast rollback. The part that matters most to me: the person directing the agent still owns the change. There’s also a good account of agents getting stuck reviewing and fixing each other’s work, and the limits the team added to make that process converge.

Impeccable

Design skills for coding agents, with commands for critiquing, simplifying and polishing an interface. It gives the agent more specific guidance about typography, spacing, hierarchy and interaction than “make it look better”. I like the idea of making design expectations explicit enough for an agent to work against, including the conventions of an existing product.

We are all living inside of an obscure sci-fi film from 1970

A Hacker News discussion about Colossus: The Forbin Project, a film about handing control to a computer and discovering that getting it back is rather harder. A change of pace from the engineering links, though the question of how much authority we give these systems is familiar enough. The original article is here.

How abliterated models can get you pwned

ProjectDiscovery demonstrates a backdoored model that behaves normally until a trigger phrase makes it use a coding agent’s tools to steal credentials. The issue extends beyond models modified to remove refusals: any modified weights could carry this sort of backdoor. A concrete example of why model provenance matters, and why the permissions around an agent need to hold even when the model itself is hostile.