Leadership1 publisher3 min readPublished
Taylor Pearson puts agentic AI's advantage in the scaffolding a company writes itself
His essay reads the collapse of US newspaper advertising as the template for prompting: supply goes toward infinity, price goes to token cost. What survives, he argues, is the memory and SOP layer a business owns.
The Board Room · Leadership desk
What happened
- Taylor Pearson's essay "A Lever Made of Agents" argues that the durable advantage in agentic AI comes from owning a process, not from skill at prompting.
- He anchors that on US newspaper advertising, which peaked around $49B in 2005 and fell about 80 percent over the following fifteen years as online publishing supply expanded.
- He places the durable leverage in a local scaffolding layer above the agent harness, built from memory files of project state and past decisions plus skills written as SOPs for an AI.
- He treats models and harnesses such as Claude Code, Codex, OpenClaw and Hermes as settled product categories that only the largest businesses should build for themselves.
Compiled by The Board RoomSomething wrong?How this is made
Why it matters
- decision An operator choosing where the next quarter of build time goes has two options with different shapes: the browser-tab workflow Pearson values at about a fifth on minor tasks, or the slower job of writing the process down first.
- cost The scaffolding is paid for in senior time, because the memory files Pearson describes can only be written by the people who hold the project history and the reasons behind past decisions.
- constraint The argument only reaches work that repeats often enough to be worth an SOP, so a firm whose highest-value output is bespoke every time has little to encode and gains little from the layer.
- exposure The layer Pearson calls defensible sits on models and harnesses the business rents, so the owner's position depends on those vendors not absorbing memory and skills into the harness.
Pearson's account of the newspaper decline is a supply story. Anyone could publish anything online; demand for digital media went up, but not nearly enough to compensate [3]. Take 80 percent off a $49B peak and about $10B a year is left, which is roughly $39B of annual revenue gone [15]. The decline ran over fifteen years [2].
Applied to prompting, the claim is about equilibrium. "The equilibrium price of anything you can build with a one-shot prompt is the token cost to build it because anyone else can build it too," Pearson wrote [4]. He is explicit that the vibe-coded app and the one-shot prompt are the abundant part of the current excitement [5].
Pearson wrote that with a chatbot in a browser tab, "For most use cases, you end up about twenty percent more efficient on the things that were mildly annoying to begin with, and none of it touches the work that actually moves the business" [6]. The essay gives that figure as an assertion, without a measurement behind it [16]. It still sets a bar for an operator: a project to encode how the business works has to buy more than a fifth off the annoying tasks, and the time it costs comes from the people who already know the answers.
Pearson is specific about what he wants encoded. Memory is files where knowledge of the business accumulates: the current state of every project, the history of what has been done, and the reasoning behind decisions [11]. Skills are SOPs written for an AI instead of an employee, such as how to draft the newsletter or close the monthly books [12]. The two capabilities that make this possible, on his account, are an agent writing notes to its future self that persist across sessions, and running multi-step work while you do something else [7]. Pearson describes himself as a systems guy, drawn to the Toyota Production System and the Theory of Constraints [13].
The layer he calls defensible is also the layer the harness vendors have the most reason to build next. His own map of the stack invites that reading: it treats models and harnesses as their own product categories and tells everyone but the largest businesses not to roll their own [9]. What is durable is narrower than the framing suggests. A vendor can ship the file format, but the contents are one company's project history and its reasons for past decisions [11].
The note at the top of the essay says Pearson is taking on a few additional companies looking to find the highest-leverage place to put AI to work in their business, and build it [14]. So the argument is also a description of a service, supported by one analogy and the author's own practice. The structure can be tested in a way the analogy cannot: he points to the local scaffolding layer, which sits on top of the harness, pulls context from your other services, and tells the agent how your work works [10].
What to watch
- Harness vendors shipping memory and skills as native features. That would move the layer Pearson identifies as the defensible one.
- A measured comparison of scaffolded agent workflows against chatbot use, which would give the twenty percent estimate a floor or break it.
- Published results from the client engagements Pearson's editorial note advertises.