Four reviews before I see the code: every change my agents build passes four levels of review before it reaches me. Structural first (does the change match the spec and the codebase's conventions), then a local code review, then cubic (an AI code-review service) reviewing the pull request in the cloud, then me. The bit that makes it work: the PR opens as a draft, so the cloud review runs and its findings get fixed before anything lands in my queue. I'm the last gate, not the first.

Finding the rhythm for a heads-down August: things go quiet in August while people are away, so I'm spending it building. The loop I've settled into: choose ticket → groom (spec it properly) → build → review. It took a fair bit of iteration on the surrounding skills to get there. The grooming bar: a ticket is ready to build only when a fresh agent session can one-shot it from the ticket body (and referenced docs) alone and verify its own work. No follow-up questions.

Cowork vs Claude Code – I changed my mind: my March post told most founders to pick Claude Code. Boris Cherny's steps-of-AI-adoption table changed my view: Cowork is the right choice for most knowledge workers – one great assistant, one task at a time. Claude Code is for going past one agent: 10, then 100, then 1,000. I've reworked the post to argue the updated version. Blog post

Ask the model to audit its own skills: I do this all the time – not a rewrite, a pass for conflicts, ambiguity and unnecessary reasoning that's crept in. Point it at the official docs first so it checks against current conventions. Tip: append .md to any Claude Code docs URL and you get the markdown version.

On my radar

Opus 5 grumbles: regression complaints everywhere this week – yet my experience is the opposite, impressed so far. Not clear why the gap, but almost certainly context-related. Same story as the Opus 4.7 release: the better the model, the less it forgives sloppy context.

Hugging Face published its intrusion timeline: the forensic follow-up to last week's sandbox-escape item. The agent spent 4.5 days inside and logged 17,600 actions, starting from a zero-day escape out of an evaluation sandbox. HF blog

Cursor Router: routes each request to the cheapest model that can handle it – Cursor claims 30–60% cost savings at equal quality. Blog post

Buzz: Block (Jack Dorsey's company) released a self-hostable, Slack-like workspace where humans and agents share the same channels – agents send patches, review work and run projects. GitHub