Think.
Whitepapers, essays, and notes from the firm. Grouped by track: what the work is about, not when it shipped.
Thesis & Proof
Can a business be built on agents? The receipts: what shipped, what it cost, what we learned.
Before You Trust AI With Real Work, Ask One Question: Who Checks It?
Every AI pitch is about volume — look how much it makes, look how fast. The part nobody demos is the part that decides whether it works: who checks the output, and whether you can trust them.
Every AI Workflow Needs Two Numbers — a North Star and a Guardrail. Almost Nobody Writes Either Down
Every AI transformation has a story. Almost none has a number. Two metrics per workflow — north star and guardrail — separate a program from a performance.
Models Write Production Code Now. Almost Nobody Built the Machine That Decides If It’s Safe to Ship
Getting a model to write your production code is a solved problem. Trusting it enough to ship is not — and almost no one has built the machine that closes the gap.
Why AI Agents Fail in Production: Four Failure Modes and the Missing Gate Behind All of Them
Most AI agents do not fail in the demo. They fail weeks later, in production, on an input no one scripted. The recurring failure modes — confidently wrong, tool misuse, silent degradation, no observability — are a measurement and gating problem, not a model problem.
The Model Picker Is a Confession: If the User Has to Choose the Model, the Product Didn’t Do Its Job
A model picker is not a feature — it is the product admitting it could not decide which model your question deserved, so it made that your job. Routing is the work, not the menu.
Generation Got Cheap. Verification Didn’t — and Whoever Closes That Gap Owns the Vertical
The literature on the AI productivity gap is measuring the wrong constraint. The cost that hasn't fallen is verification. Until it does, the gradient widens — and whoever closes it captures the vertical.
Talent Density Is the New Headcount: Why the Smaller, Better Team Now Beats the Bigger Budget
Two attempts at powered flight, two months apart, in 1903. The team with the bigger budget crashed into the Potomac. The team that ran a bicycle shop ended the century.
Forward-Deployed Engineer Hiring Is Up 1,000%. That’s a Confession About What AI Products Can’t Do Alone
Forward-deployed engineer postings at the AI labs are up more than 1,000 percent year-over-year. The motion says more about what the products cannot yet do than about what enterprises cannot yet absorb.
For a Brief Window, You Can See Exactly What a Thought Costs. The Window Is Closing
For a brief window, we can see the meter on the imitation of cognition. Either someone clever hides it, or abundance dissolves it. Either way the window closes.
Information Asymmetry Has a Five-Stage Arc — and AI Just Triggered Stage 5
Why thirty years of leveling the playing field failed, and what's finally changing. A five-stage framework, with citations from Akerlof to the AI Index.
A Letter From the Founder: Why A8C Builds Where Information Asymmetry Costs People the Most
The asymmetry that costs ordinary people and small businesses isn't going away. But for the first time, it doesn't have to stay decisive. A note on what we're building, and why now.
Agent Infrastructure
Research from building the OS that builds the calculators — the colleague pattern, model economics, what breaks at scale.
Before You Trust AI With Real Work, Ask One Question: Who Checks It?
Every AI pitch is about volume — look how much it makes, look how fast. The part nobody demos is the part that decides whether it works: who checks the output, and whether you can trust them.
Every AI Workflow Needs Two Numbers — a North Star and a Guardrail. Almost Nobody Writes Either Down
Every AI transformation has a story. Almost none has a number. Two metrics per workflow — north star and guardrail — separate a program from a performance.
Models Write Production Code Now. Almost Nobody Built the Machine That Decides If It’s Safe to Ship
Getting a model to write your production code is a solved problem. Trusting it enough to ship is not — and almost no one has built the machine that closes the gap.
Why AI Agents Fail in Production: Four Failure Modes and the Missing Gate Behind All of Them
Most AI agents do not fail in the demo. They fail weeks later, in production, on an input no one scripted. The recurring failure modes — confidently wrong, tool misuse, silent degradation, no observability — are a measurement and gating problem, not a model problem.
Forward-Deployed Engineer Hiring Is Up 1,000%. That’s a Confession About What AI Products Can’t Do Alone
Forward-deployed engineer postings at the AI labs are up more than 1,000 percent year-over-year. The motion says more about what the products cannot yet do than about what enterprises cannot yet absorb.
For a Brief Window, You Can See Exactly What a Thought Costs. The Window Is Closing
For a brief window, we can see the meter on the imitation of cognition. Either someone clever hides it, or abundance dissolves it. Either way the window closes.
No, MCP Is Not the New SEO: What the “Agents Are the New Buyers” Take Gets Wrong
A viral take says AI agents are the new buyers of the internet, and any business without an MCP server will be invisible to them. The analogy fails, and the failure tells us what to build instead.
AI Labs Are Optimizing for Calculator Skills — and Selecting Against What Makes a Colleague Valuable
The AI industry is optimizing for calculator-properties — confident, fluent, accurate-on-the-test. The properties that make a colleague valuable are exactly the ones that pressure selects against. So someone has to build the colleague anyway. Here's how the studio model gets the work funded.
AI Labs Say Hallucination Is a Phase Models Grow Out Of. The Analogy Collapses in Two Places
AI labs say LLMs hallucinate the way toddlers confabulate — and grow out of it the same way. The analogy collapses in two places, and where it collapses tells you what's actually being built.
Vertical Intelligence
Domain knowledge that emerges from each calculator — what the data actually says about the industries we build for.
The Club Coach Runs a Seven-Figure Business on Group Chats: The Most Under-Tooled CEO in America
The club coach runs a seven-figure operation on group chats and gut — in a market where the families funding everything know the least.
No, MCP Is Not the New SEO: What the “Agents Are the New Buyers” Take Gets Wrong
A viral take says AI agents are the new buyers of the internet, and any business without an MCP server will be invisible to them. The analogy fails, and the failure tells us what to build instead.
AI Labs Are Optimizing for Calculator Skills — and Selecting Against What Makes a Colleague Valuable
The AI industry is optimizing for calculator-properties — confident, fluent, accurate-on-the-test. The properties that make a colleague valuable are exactly the ones that pressure selects against. So someone has to build the colleague anyway. Here's how the studio model gets the work funded.
Most Prestigious Knowledge Work Is a Scribe Job With Better Marketing — AI Just Made It Obvious
Most prestigious knowledge work is a scribe job with better marketing. AI just made it obvious.
Information Asymmetry Has a Five-Stage Arc — and AI Just Triggered Stage 5
Why thirty years of leveling the playing field failed, and what's finally changing. A five-stage framework, with citations from Akerlof to the AI Index.