One thing for me, many for the agents. The classical focus methods (Ivy Lee's six tasks, do-one-thing, etc) and working with parallel agents seem like they contradict. They don't. Your attention is scarce and protected. Agent attention is abundant and orchestrated. The trap is using agent leverage to take on more deep work for yourself instead of expanding the number of parallel streams the agents run. Freed-up time should go to spec-writing and review capacity, not new foreground commitments.

Three gates before any task goes background. 1. Spec-complete: could a competent colleague execute this with no follow-up questions? 2. Low iteration cadence: can it complete the task in 1–3 messages or does it need lots of back-and-forth? 3. Self-verifiable: does the agent have a way to know it's done? Fail any → keep interactive.

Investment order matters more than parallelism count. Most 'scale to more agents' attempts skip the lower rungs. The order that actually compounds: verification capacity → codebase legibility → environment context → spec quality → skill library → layered review. The bottleneck migrates as you move up. Early on it's "how do I know it works?" Then "the agent keeps adding the wrong pattern", then "the agent built the wrong thing", and eventually "I have 8 PRs and no time to look at them". Adding capacity to a layer that isn’t the bottleneck doesn’t increase velocity.

/simplify is now /code-review. Same skill, renamed. Worth knowing the new args: /code-review high runs at higher effort, and --comment posts findings as inline GitHub PR comments instead of returning them to your terminal. Tweet

Skills are how operating principles survive a busy week. Built two new skills this week, both codifying the cadence above: an end-of-day close ritual and a weekly review. The discipline behind both is the same question: why am I doing this manually? Whatever you stop to do twice is a skill waiting to be written.

On my radar

Open-source model guardrails come off in under 10 minutes. A technique called abliteration has produced 3.5k+ uncensored variants of Llama and Gemma, downloaded 13 million times. Proprietary models like Claude can't be touched as the weights aren't accessible. The interesting bit is the regulatory tension between open-source norms and safety enforcement. Hugging Face example

An OpenAI model disproved a 50-year-old geometry conjecture. Cross-domain proof keeps stacking up. It's being called a 'Move 37 moment' – the AlphaGo term for the move no human would have played that turned out to be the right one. OpenAI

Anthropic is heading into its first profitable quarter. $350m of new ARR per day, growing faster than the infrastructure can keep up with. Context for every outage and rate limit complaint. WSJ