Writings

Technical

• AI agents need millions of computers. Do they actually need to? Oct 10, 2026

  Every training attempt gets its own computer, and they are all nearly identical. A walk down the stack: what snapshots and copy-on-write already share, what is still duplicated, and what CUA-Sandbox found when it shared the running software itself.

• What does it cost to give an AI one more try at using a computer? Oct 9, 2026

  Training an AI to use a computer means giving it a computer to practice on, hundreds of thousands of times. I read four papers to find out what one attempt costs in time, tokens, and machines, and where it all goes.

• I hacked Microsoft Oct 6, 2026

  Well, almost. The code Microsoft Research uses to grade AI agents ran eval() on text the agent could control, so the thing being tested could run code on the thing testing it. Plus a second bug that quietly scores finished tasks as zero.

• Human-level on the benchmark, can’t click OK on a MacBook Oct 4, 2026

  Two fixes to Agent S, the open-source computer-use agent: a screenshot too big for the model server to accept, and clicks that landed 300 pixels off on every screen that wasn't the benchmark's. What a 66% on OSWorld doesn't tell you.

• Three bugs where the tool said "done" and nothing happened Oct 4, 2026

  Three open-source fixes in a week to computer-use tooling: an Enter key that never arrived, a click that froze on a confirm() dialog, and an iframe that stole the address bar. Same bug, three costumes.

• How to fork a computer Sep 30, 2026

  Freeze a running machine at one exact moment and wake up two of them. How Firecracker, VMware, QEMU, and CRIU do it, why it's fast, and why the children wake up confused about who they are.

• The three layers of agent persistence Sep 30, 2026

  A persistent agent needs its thinking, its computer, and its logins to all survive. Lose one and the other two are mostly useless. Plus VMs, containers, Firecracker, and why everyone is suddenly talking about forking.

• What is computer use, actually? Sep 24, 2026

  A computer-use agent is three separate things pretending to be one: a model, a loop, and a computer. Pull them apart and the whole space gets easier to think about.

• Everyone is building the same dashboard Sep 22, 2026

  A/B testing, analytics, error tracking, and observability were four tools. PostHog, Sentry, Datadog, and Amplitude are converging on one, and the feature flag is what's pulling them together.

• We Stopped Building Everything: Abstraction Ate Software Architecture Sep 13, 2026

  With Rahul. Booch says software history is rising abstraction. The architect's job went from building systems to assembling them, and Garlan called it in 2000.

• What actually happens between writing code and running it? Jul 4, 2026

  I asked why Rust needs a linter if it's type safe. A few questions later, compiled versus interpreted stopped being a real category. A six part series, from your keyboard down to the CPU.

• Nobody can review a hundred agents Jul 2, 2026

  Every agent tool is racing to run more agents. The bottleneck was never writing the code, it's the human who has to review it.

• My agents die when I close my laptop Jul 1, 2026

  Coding agents are sold as tireless workers, but they're processes on a laptop that sleeps. On persistence, SSH, and haunted printers.

• My coding agents can't see my whole product Jun 30, 2026

  Every orchestrator assumes one task lives in one repo. My product is two repos, and the workaround is a fake parent git repo.

• I tried to prove AI gets dumber the more you talk to it. It refused. Jun 18, 2026

  A 12,570-trial hunt for "context rot" that came back empty - and the scariest bug was in my own scorer.