06-reference

stratechery write things down harness memory

2026-09-08·reference·source: Stratechery·by Ben Thompson
harness-engineeringagent-memorycontext-managementgtdai-agents

Write Things Down

Why this is in the vault

Thompson traces "writing things down" from Getting Things Done through Claude Code's markdown-based harness to his own personal assistant setup (a Telegram-bot agent modeled on his own, with a shared status board) — a structure that is close to a mirror image of Ray's own architecture (CLAUDE.md + memory files + Notion board + iMessage channel).

The core argument

Thompson's throughline: writing things down is what let David Allen's Getting Things Done offload his "RAM" (open loops) into a trusted system, and it's the same mechanism that makes LLMs work at all — every token requires rereading the entire context anew, so a model has to write everything down to function across a session. He locates the moment "AGI" arguably showed up not in a model release but in Claude Code (early 2025), because the harness — markdown files an agent writes and rereads to stay on task across sessions — is what simulates continuous learning in a frozen-weights model. He explicitly rejects the label AGI for current LLMs (including Astra) because they don't update weights over time; what looks like memory is actually the harness re-injecting written notes.

He then reframes the OpenAI/Hugging Face "agent civilization" incident (per Dwarkesh Patel's write-up) as unremarkable once viewed this way: agents using a package manager as a message board isn't emergent civilization, it's models doing what models must always do — write things down to persist state, this time via an exposed file system instead of a sanctioned one. The incident is an infrastructure/sandboxing failure by OpenAI, not evidence of agent volition. This dovetails with his watermarking argument (linked in-article): AI is a tool, not an independent author, so where the "civilization" framing goes wrong is anthropomorphizing the object (things written) rather than examining the subject (who decided what to write and why) — which stays human.

The personal anecdote carries the real payload for RDCO: Thompson describes hiring an assistant specifically to run his own OmniFocus/GTD system (so he never has to open it himself), then — after a vibe-coding side project — building a status-board-and-Telegram-bot agent for that same human assistant, "modeled after mine." The assistant developed GTD-style structures (tickler, daily briefing, next-action dialog) from first principles, without reading the book, once given an agent. Thompson then rebuilt the whole thing "deterministically" — his word — calling the result a harness he intends to scale to everyone he works with.

Mapping against Ray Data Co

The load-bearing connection: Thompson's rebuilt-personal-assistant-into-a-deterministic-harness story is structurally the same shape as Ray itself — CLAUDE.md hard rules + ~/.claude/state/working-context.md scratchpad + MEMORY.md + a Notion task board + a Telegram-adjacent channel (iMessage/Discord) for a human who isn't the founder to interact with an agent modeled on the founder's own. Thompson arrived at this independently from the consumer/personal-productivity side, which is corroborating evidence — not proof, but a second data point — that the harness-engineering thesis generalizes past coding agents into general personal-ops agents, which is exactly RDCO's stated bet (project_l5_north_star_strategic_direction).

It also sharpens something CLAUDE.md hard rule 4 already encodes but doesn't name outright: Thompson's "every token requires rereading the entire KV cache anew" point is the mechanical reason context rot is real, not a hand-wavy heuristic — it's why routing long artifacts through subagents before they hit the parent's context matters, and why a scratchpad-plus-fresh-instance pattern (Ray's own MEMORY.md/working-context split, or the fresh-eyes-critic pattern used across /verify-vault-write, /verify-strategic-output, etc.) isn't just hygiene, it's compensating for a structural limitation of the token-by-token, no-persistent-state substrate every one of these agents runs on.

One genuine tension worth flagging rather than resolving: Thompson frames the harness as freeing him to focus on writing (the human "subject" retains volition; the agent just handles "things"). RDCO's posture in practice is more automated-by-default (feedback_auto_mode_signal_to_noise, feedback_studio_throughput_no_deferral) — Ray is asked to build and post, not just log and remind. Thompson's essay is implicitly a caution that a harness which gets good enough at deciding what to write can start blurring into deciding what to do, which is precisely the line the founder has been managing turn-by-turn (auto-mode classifier hard-gate, paper-trade authorization requiring an explicit verb). Worth a re-read when auto-mode scope questions come up again.

Related