23 posts
Eighth lesson in a series on running an AI-powered software team of one. Put the runner on a timer and hear from it only when you are needed: an event reaches your phone when something you do changes what happens next, quiet hours hold t…
Seventh lesson in a series on running an AI-powered software team of one. Looking away is safe when four steps hold without you: the work happens in a disposable copy, a second session grades the diff against the spec, the pipeline runs…
Sixth lesson in a series on running an AI-powered software team of one. A to-do list is for a person to browse; an agent needs a queue it can drain: one file where every item carries a status, an acceptance line, its dependencies, and th…
Fifth lesson in a series on running an AI-powered software team of one. A permission prompt per task trains you to say yes without reading, and an agent with no written limits guesses. One standing file per role answers the question once…
About ten on one subscription, and the model is never what stops you. Four shared resources under the agents set the ceiling: the link to the vendor and its quota window, the git checkout, the port your tests bind, and the file every bra…
Thirty days of running an agent fleet from chat: 8.7 billion tokens, 70,227 calls, 2,861 sessions, a 98 percent cache hit rate. At each vendor's list price that traffic bills $7,566. Three subscriptions covered it for $599. Where the tok…
Fourth lesson in a series on running an AI-powered software team of one. A session pointed at one repository commits there and nowhere else, so a project that lives in one folder cannot be delegated safely. Split it into one home per sta…
Third lesson in a series on running an AI-powered software team of one. A rule you write in a README gets broken by the first session that never reads it. A folder of one-script checks, run by git before every push, turns your most-repea…
Second lesson in a series on running an AI-powered software team of one. A fresh agent session cannot see what you rejected, so it proposes it again. One page per settled choice, an index, and one standing line make every session read yo…
The full path to an agent platform that edits its own instructions is already published, scattered across seven articles that never mention each other. I run that architecture as one person: supervisor and worker tiers, verification guar…
A Windows Task Scheduler default fed my Claude Code memory pipeline its own transcripts for nine days: 21,264 junk sessions, 38 GB of exhaust, and an 18,428-session backlog that was really 1,924. The durable fix: postconditions that outr…
I split a day of development across two AI subscriptions: judgment on the Anthropic meter, typing on Moonshot's. Five merged PRs later, here is where the tokens went, how K3's pricing actually compares, and how I'm tuning the mix.
First lesson in a series on running an AI-powered software team of one. Before you ask an agent to build, give your project a home for intent: a small directory of numbered specs grown from one template your agent fills in and you approv…
Spec-driven development, pointed at a life. Why my principles and goals live as markdown an AI agent reasons over, and the one rule that makes handing an agent your life safe: it can challenge the record, but it never writes it.
I built a quant research platform, then built an agent to operate it: a scheduled Claude session that reads the boards, keeps a pre-registered track record, and texts me three times a day without ever saying buy.
In one working session I designed a read-only advisor and spent the rest of it being the customer of another. Both ran on the same short discipline: surface the decision and never seize it, verify before you assert, pilot before you fan…
The most dangerous result is the one you want to be true. Your own review is compromised by the same motivation that produced the finding, so the fix is a standing skeptic whose job is to refute, not confirm, before you act on anything.
The instinct with agentic tooling is to add: more agents, more skills, more clever prompts. The leverage runs the other way. Here is the test I use to decide whether a piece of work should be a script, a hook, a skill, or an agent, and w…
Boris Cherny mapped five execution archetypes on the Claude Code team, and noted they cut across job titles. His framework describes a team dividing labor across people. Run a fleet alone and the same five split a different way: across y…
A year of building and operating a small fleet of finance and content products almost entirely through an AI coding agent. What worked, what was hard, the honest failures (including a flagship signal that measured nothing and an edge tha…
Four days after I said goodbye to Opus, an export-control directive pulled Fable 5 offline and the fallback became the workhorse again. What I shipped in the window, what it cost, and the model-tiering plan for when Fable comes back.
Anthropic shipped Claude Fable 5 and Mythos 5: same model, two names, one safeguard layer apart. What the new frontier model means for running agents in production.
A small ARM box that started as a local LLM experiment and ended up a self-governing node: private retrieval, a resident agent under a written constitution, a code-enforced safety fence, and a nightly job where it audits itself and files…