Field Notes

Thinking, in the open.

A garden, not a blog. It opens with what I'm doing right now, then the thinking underneath it. Some of this is finished and I'll argue for it. Some is raw and still moving. I'd rather show the work than wait for it to be perfect.

Now

The live snapshot: what has my attention this season. Everything below it is the thinking that comes out of it.

Building
  • This site, treated as a real product. The garden below just woke back up with four new entries.
  • A fleet of agents that runs my daily work, with review gates enforced by a model from a different family than the one doing the work.
  • An AI school with one student, me, that won't release the next lesson until I teach the current one back.
Learning
  • The Anthropic certification track, delivered through that school. Currently in the engine room: what the API actually is, and what an agent actually is.
  • Graph engineering, the emerging layer above prompt and context engineering.
Evaluating
  • Buzz, Block's new team platform where AI agents sit in the channel as members. Day-one note is in the garden below.
Pointed at
  • A forward-deployed engineering role, building for a real customer instead of only being my own.

Last updated August 2026. This section is a now page, living at the top of the garden.

Essays

Finished thinking. Things I've worked out well enough to argue for.

5 minThe school that has to win its student backI built an AI curriculum that won't release the next lesson until I teach the last one back. Then it went silent for twenty days, and the silence taught me more than the lessons did.educationagentsautonomy5 minWhat changes when one agent becomes a fleetMore agents doesn't mean more done. It means your judgment has to move up a level.multi-agentorchestrationreview6 minPersistent identity in multi-agent systemsAn agent that forgets at the end of every session isn't an assistant. It's a stranger who happens to be good at your job.multi-agentmemoryarchitecture5 minWhy mechanical gates beat soft disciplineYou can't prompt your way to safety. The rule that matters is the one the system can't talk itself out of.agent-safetygovernanceharness

Notes

Raw and in progress. Working notes from things I'm still in the middle of.

3 minFirst look: Buzz, where the agents sit in the channelJack Dorsey's new team app treats AI agents as members, not features. Day-one notes from the start of an evaluation.evaluationtoolsmulti-agent3 minWhat a year of running agents actually costsThe honest, unglamorous ledger nobody puts in the launch post.operationscostreality-check

Patterns

Shapes I keep reaching for. Small, reusable, hard-won.

3 minOne pane of glass for the whole fleetWhen every agent keeps its own diary, the system's status lives in nine places, which means it lives nowhere.patternobservabilityoperations2 minAppend-only state as the unit of continuityWhen agents hand work across sessions, make the log a ledger, not a whiteboard.patternmemorystate