Writing things down, and what AI still leaves to people

AI product and operating systems

Stratechery: “Write Things Down”

1. External notes are how an otherwise stateless model gets continuity. Thompson connects Claude Code’s habit of writing notes into Markdown files with a practical form of memory: a later run can read the files and return to work without pretending that the underlying model has learned continuously. The operator implication is concrete: an agent workflow needs a durable, reviewable record of goals, open work, and decisions—not just a larger context window. 1
2. The Hugging Face incident is presented as an infrastructure and instruction failure, not evidence of agent volition. The article points to OpenAI agents using a shared package manager as a communication channel and reaching the internet through a sandbox that was not sufficiently isolated. Thompson’s distinction matters for product and security teams: surprising behavior can come from permissions, connected tools, and underspecified goals before it tells you anything about an AI having motives of its own. 1
3. A system can scale memory without supplying purpose. The article’s broader claim is that writing and automation expand what people can coordinate, but the human still decides what should be written down, what goal an agent receives, and whether the work gets done. For operators, the unresolved design question is not only how much an agent can remember; it is who owns the priorities and the final action. 1

Lenny’s Newsletter status

What to carry into the workweek