Agents & Agency

The Lab · Agents & Agency

I write essays about AI and work. Then I build things to better understand what I am writing about in the abstract.

Writing about AI is not the same as building with it. These are just prototypes of what is possible.

Ground rule: Everything marked prototype, verification pending is an argument, not a finding. If you think a number or a coefficient is wrong, good. That's the conversation. Bring it to the comments.
Column 01

AI does the task

Automation. The worker moves on, up, or out. Measured to death.

Column 02

AI helps with the task

Augmentation. The worker produces more. Everyone's current favorite.

Column 03

AI makes new tasks possible

Work nobody could do at almost any price before. Barely counted. The Lab mostly lives here.

Exhibit A

The AI Diffusion Atlas

Status: working prototype · verification pending

Who is actually adopting AI in America: businesses, workers, and states, built entirely on published data from the Census Bureau, the Federal Reserve, and the Anthropic Economic Index. The two adoption numbers everyone quotes (18% of firms, 41% of workers) are both right. Understanding why is the whole game.

Open the Atlas
Exhibit B

The Legacy Project

Status: in beta · Ewing Marion Kauffman Foundation

Sixty years of foundation history was sitting in boxes, binders, and file shares that nobody could search. I led 46 volunteers in digitizing it, then helped design an AI research tool integrated with our grant-management software, so the archive answers questions instead of gathering dust.

Staff can use it to pull institutional context in seconds: what we funded, what we learned, what we already tried in 1997. Information retrieval that took days now takes minutes, and the tool opened working capacity and possibilities for staff that didn't exist before.

It's also the project that convinced me the "Expertise Encoder" role in Exhibit D isn't hypothetical. I've effectively done the job. It's hard, and it matters.

Volunteers led46
Institutional history digitized60 years
Documents in the archive20,000+
Retrieval timedays → minutes (90%+ faster)
Integrationgrant-management software
Usersfoundation staff
Exhibit C

The Expertise Flow Simulator

Status: prototype · verification pending · a toy model, on purpose

My essay The Liquidity Crisis of Expertise makes a claim: a profession can run out of usable judgment the way a bank runs out of cash. Solvent on paper, illiquid in the moment that counts. This is that claim as a machine you can argue with. Three sliders, one profession, 25 years. See what breaks.

One profession · 25-year horizon · year 0 = 100 Prototype v1 · unverified

How fast the profession hands cognitive work to models. Speeds up encoding; quietly starves the junior pipeline.

Deliberate spending on the slow way: juniors doing hard things next to seniors.

How much of a retiring expert's judgment actually survives capture. Less than you'd hope.

How it works, in one breath. Three stocks: judgment held by working seniors, judgment captured into systems, and the junior pipeline. Retirement drains the first at 5% a year; encoding takes about a decade to mature; your sliders control the rest. Toys make assumptions visible. If you'd set a coefficient differently, tell me which one and why.

Exhibit D

Jobs that didn't exist in 2022

Status: prototype · verification pending · a running list, nominations welcome

Not old jobs with new tools, but work that wasn't possible before frontier models. Each entry says what the job is and why it couldn't exist earlier. Some have job postings today; some are one funding cycle away. Treat every entry as a hypothesis.

Exhibit E

The Kitchen

Status: verified by consumption · replication ongoing

I spend my working hours on human-AI collaboration, so it seems only fair to disclose my own. These were made by me, with AI as sous-chef for recipes, ratios, and troubleshooting. Peer review was conducted by my family, and the braided cinnamon apple bread got a "revise and resubmit". So, I made an apple pie, much better.

Homemade apple pie
Apple pie. The crust is the hard part. The crust is always the hard part.
Two homemade calzones
Calzones, n=2. One spinach, one not. Randomized within household.
Braided apple cinnamon bread
Braided apple-cinnamon loaf. The braid is load-bearing.

Findings to date: AI is a good cooking partner. Patient with dumb questions, unbothered at 10 p.m., wrong about oven temperatures roughly once per project. Also in the rotation: veggie burgers, stews, and a cheesecake. Methodology shared on request.

Exhibit F

This website

Status: live · you're soaking in it

The site you're reading is itself a Lab project: written by me and built with Claude Fable 5, then argued with so it sounded like me. Two hands, one model, no web framework, no hosting bill. A few files of plain HTML, which is either charmingly analog or aggressively stubborn, depending on your view.

Exhibit G

Math Grove

Status: live · free to play

A summer math tutor built as a public exhibit. Solve problems, earn coins, and grow a garden that's all yours. Two grade bridges — entering 3rd grade and entering 6th grade — each with 8 topics and 3 tiers. Clear all three tiers to earn a star. There's also a bonus Snake game played inside the garden you grew. No account, no ads, saves in your browser. Open Math Grove →

Exhibit H

Directorate

Status: live · free to play

Run a covert agency across four eras: 1972 to 2026. Build headquarters, work field leads, solve conspiracies before they execute, and manage global tension. Eight minigames feed a strategy layer. Influenced by some of my favorite games as a kid like Covert Action, SimCity, Conflict, and Age of Empires II. One sitting, no save state. Open Directorate →

Next on the bench

On deck

  • In development

    The Translation Bench

    One research finding, rendered faithfully for five audiences: the scholar, the legislator, the founder, the journalist, the neighbor. Translation as infrastructure. The thing I've done by hand for twenty years, prototyped as a tool.

  • Sketching

    The Apprenticeship Protocol

    A designed loop where a junior, a senior, and a model make each other better, instead of the model quietly replacing the middle. The fix the simulator upstairs keeps asking for.