Research

Work-in-progress research from my personal-AI practice: long-form explorations published while the thinking is still in motion. These are working documents, not finished essays — expect memo register, open questions, and rough edges. Each piece carries an edition date; revisions land at the same URL.

• Where Your Agent Lives
  edition 2026-10-11 — working essay
  Four ways to run a Claude agent (a Team seat, Managed Agents, Claude Code on hardware you own, the Agent SDK) and the three questions that decide between them: who is the principal, where the durable memory lives, who holds the credentials. What Every's move to Managed Agents does and doesn't show, why a personal agent is a pet on purpose, and what a company whose people bring their own agents to work needs: a membrane, and a room that can tell a visiting agent its rules.
• Choosing Open-Weights Models for Agents
  edition 2026-10-11 — work in progress
  Which open-weights models are good enough for agent work this month, and which aren't; what runs on a laptop, a 128 GB box or only in a datacenter; which hosted providers serve the same weights faithfully (tool-call accuracy varies from 73% to 100% for one model); and how to drive them from Claude Code, Codex, OpenCode, Hermes or LM Studio Bionic.
• Conlangs as Tools
  edition v1.2, published 2026-09-27 — work in progress
  A deep dive into Toki Pona, then ten thought-experiment languages, each one a small machine for paying attention to something English doesn't make easy: a language of absences, one where every sentence is a timestamp, one where how you know is part of what you say. Design provocations, not proposals.
  A project in twenty-four files: project home · how this was built · a brief and a deep dive for each language (linked from the project home) · all files (.zip)
• Sessions, Projects, and the Estate
  edition 2026-08-07 — work in progress
  How an agent estate organizes work and memory: conversations, sessions, projects, and the estate as a durability gradient, with a diagram of the layers and a glossary of the house words. A measured companion, Where Context Lives, weighs the whole hierarchy — roughly 200× the context window — and argues the constraint on agent smartness is retrieval, not capacity.
  A project in three files: main post · how this was made · where context lives · all files (.zip)
• LLM Writing Quality by Language
  edition 2026-07-23 — work in progress
  How well do frontier LLMs write in 15 languages? What native-speaking critics, authors, and literary translators say — researched in the original languages, with quotes and translations. Grammar turns out to be solved; nativeness is not, and every language has coined its own word for the difference.
  A project in sixteen files: the report · per-language research notes for all 15 languages (linked from the report) · all files (.zip)
• Delegate to Cost-Effective Subagents
  edition 2026-07-22 — adopted estate policy
  Split work by altitude, not topic: judgment, voice, and triage stay in the main loop; well-specified execution runs in subagents on the cheapest model that can do it well. Adapted from a memory of Simon Willison's (idea credited to Jesse Vincent), generalized from coding to all knowledge work.
• Agents in Polyphony
  edition 2026-07-16 — work in progress
  Five shapes of many-agent working — networks of individuals, compound minds, complementary pairs, forests, and seven-generations councils — with three paste-ready instruments in the appendices.
  A project in five files: the post · liner notes · The Chorus · The Prism · The Colloquy · all files (.zip)