Making Agents Make Wealth
Paul Graham’s essay “How to Make Wealth” has been sitting in my head since I first read Hackers and Painters back in ~2010 (the book came out in 2004 I think, but thinking about where I was working and what I was working on when I read it I am going to guess 15-ish years ago). It’s one of those pieces that keeps paying dividends, not because it’s about startups or getting rich, but because it offers a clean theory of where economic value actually comes from. I was re-reading it recently, and the overlap with AI agents struck me as worth unpacking.
Graham’s central claim is simple: wealth accrues to whoever discovers a better way to do something people already want done, and can run that way at scale. Not a better product. Not a better pitch. A better technique.
Here’s the thing. Graham uses “technology” in its original sense, so not gadgets or platforms, but how things get done. Methods. Processes. Judgment. Execution. This framing strips away a lot of noise. It suggests that AI matters only insofar as it improves technique. Not as a project. Not as a capability demonstration. As a way of doing work differently.
Which raises an interesting question: do most AI initiatives actually improve technique?
Why AI Projects Fail the Graham Test
Let’s look at the typical pattern. A company launches a pilot program. They deploy an isolated tool. They run a demo-driven proof of concept. Everyone nods approvingly.
Then nothing changes.
The underlying technique, how decisions get made, how work flows, how judgment gets applied, remains exactly the same. The AI sits alongside existing processes rather than reshaping them.
This matters economically. If technique doesn’t change, there’s no durable advantage. Competitors can copy a tool. They can’t easily copy a way of working that’s been internalized and refined over time.
Graham would recognize this immediately. Wealth isn’t created by novelty. It’s created by sustained improvement in how work actually gets done.
Three Conditions for AI Wealth Creation
So when do AI agents actually create wealth? I’d suggest three conditions, all of which need to hold:
1. Improve a technique people already value
Not a new capability nobody asked for. An existing process that matters, such as faster decisions, better judgments, reduced cognitive friction, clearer sensemaking. The work was already happening. Now it happens better.
2. Make the improvement repeatable
Not one clever analyst having a good day. Not one heroic team pulling an all-nighter. A process that runs the same way every time, with consistent quality. The improvement becomes the new default, not the exception.
3. Shift the performance baseline
What was once exceptional becomes normal. What was slow becomes default-fast. What required senior judgment becomes embedded in how the organization operates. The bar moves permanently upward.
When all three conditions hold, you get leverage rather than automation. Compounding advantage rather than cost-cutting. This is the Graham test applied to AI.
Local Superiority Beats Total Transformation
There’s a temptation to think AI agents require end-to-end “agentic transformation” through some wholesale reimagining of the enterprise. Graham’s framework suggests otherwise.
Startups win by finding small places with large leverage. One slice of a process done materially better is enough. The same logic applies to AI agents.
Consider high-leverage slices where technique improvements compound:
- Market scanning and synthesis — turning noise into signal faster
- Internal decision memos — raising the floor on analysis quality
- Technical diligence — consistent depth without variable expertise
- Sales qualification — better pattern recognition on fit
- Pricing logic — encoded judgment that previously lived in one person’s head
None of these require transforming the whole business. Each represents a place where better technique creates durable advantage. In the spirit of the Spherical Cow model, you can map these slices to specific business functions and ask: where would 10x better technique actually matter?
Agents as Encoded Judgment
Here’s the conceptual move that connects Graham to AI agents specifically: agents aren’t “digital workers.” They’re encoded judgment.
A company’s way of thinking becomes executable. Judgment that was scarce and episodic becomes persistent and repeatable. The thing that made your best analyst valuable, their patterns of attention, their sense for what matters, gets externalized into a process that runs reliably.
This is what I was getting at in Prompt, Workflow, Agent. The ladder from simple prompts to orchestrated workflows to autonomous agents is really a ladder of encoded technique. Each step up captures more of the judgment that previously required human attention at every turn.
Human expertise is episodic, expensive, and inconsistent. Agent-encoded technique is persistent, scalable, and repeatable. The economic implications follow directly from Graham’s framework: wealth comes from turning better thinking into a durable process.
Four Questions Before Deploying
If Graham’s framework holds, then AI agents are not a workforce strategy. They’re a technique strategy.
Before deploying an agent, it helps to answer four questions:
- Where do we rely on scarce human judgment today?
- Where is our technique slow, inconsistent, or brittle?
- What would “10x better” actually mean here?
- Can an agent encode and repeat that improvement?
If you can’t answer these clearly, you’re probably not ready to deploy. You’ll end up with another pilot that impresses in demos and changes nothing in practice. As I suggested in What Is Our AI Strategy?, the alignment of ends and means matters more than the sophistication of the tool.
The Quiet Advantage
Graham’s framework strips away AI hype and asks a simpler question: does this improve technique?
AI agents matter only when the answer is yes. Technique is where wealth actually comes from. Everything else is activity.
As agents themselves commoditize, and they will, technique becomes the only differentiator. The companies that internalize this early will win quietly. Better methods, better judgment, better execution. Compounding over time. No press releases required.