Founding Fathers of Agents

Three thoughts I kept coming back to, until I realized they were the same thought.

One. Values are simple. Scaling them is not.

A friend in San Francisco recently said something that’s stuck with me for weeks.

The moral and social opinions we hold about how people should treat each other — what’s fair, what’s owed, what’s forbidden — truly are political stances in disguise. And there is nothing wrong with thinking this way, since it shows up immediately upon generalizing these opinions to apply across a huge, imperfect population, over hundreds of years. Politics is the friction you get when a moral idea meets a hundred million people and a hundred years.

At the other end of that spectrum sit the constitution, the body of law, and the long-built habits of government. Together they form a careful system of checks and balances. The whole point of that system is that one bad decision — even one the majority loves! — can’t ruin the entire country. The constitution isn’t a statement of values. It’s the harness that lets imperfect people carry values across generations without burning the place down.

That changed how I think. I used to treat my own moral and social opinions the way most people do — as standalone claims to defend or fine-tune. “This is right. That is wrong. Here’s why.” But that’s the weak, self-comforting framing. The stronger framing is: how does this rule of mine grow into a system that lasts? From Kant’s categorical imperative, which is moral and principled, to the pragmatist approach of “what works, at scale?”

A system that survives time, opponents, idiots, and future versions of me who may have forgotten why the rule existed in the first place. A rule that only works when good people enforce it isn’t a rule. It’s a wish.

I respected constitutions before, in a vague civics-class way. What I didn’t fully grasp is how enormously hard the work of writing one is. Crafting a document that turns a population of selfish, often uninformed, often emotional, mostly uneducated and opinionated people into a society that creates lasting value across generations — that’s one of the hardest design problems humans have ever tried. Most attempts fail. The ones that succeed deserve far more credit than they get.

So my first realization is small but real: I want to take moral and social opinions — mine and other people’s — more seriously, and judge them through a harder lens. Not “is this the obvious answer?” or “does it feel right to me?”, but “what kind of system does this turn into when you scale it across imperfect people and long stretches of time”, and “does that system survive a century of stress?”

Two. Agentic systems are the same, just faster.

Switching gears, today I am writing this under the assumption that most of my readers are engineers. By now it’s clear that as of 2026, the biggest area where generative AI creates real economic value is coding. The reasons are obvious: very fast and reliable feedback loops — both human-in-the-loop and machine-in-the-loop — and huge downstream value from writing good code faster and more reliably.

So everyone and their uncle is now building agentic systems that produce code — or things that are basically code. Rules, policies, procedures, the data-shuffling logic for pulling from a spreadsheet, checking something on the web, and dropping a row onto your calendar. You might not see a diff, but behind the scenes it’s still code. The bar for an agent to write the connector glue is so low now that you might as well imagine all of it happening inside your coding assistant: hand it tokens to read and write your Spreadsheets and Docs and to access your calendar, and let it do the rest.

The popular framing right now is that every senior-plus engineer has become a small tech lead managing their own swarm of agents. I don’t think that’s quite right. The more honest framing is that engineers are becoming founders of small tribes of agents.

The job of us as tribe leaders is not to hand out tickets to our “direct reports”. Our job is to design a meta-harness — really a constitution! — that takes a goal and turns it into coordinated agent work whose sum is greater than its parts.

And here’s the thing: that isn’t really an engineering problem. It’s the same scaling problem from the previous part of this post, just with a different population and a faster clock. The failure modes rhyme: both agentic harnesses and failed states tip toward anarchy on one end or over-centralized rigidity on the other. The challenge is to build something between those two poles that actually produces coordinated value over time, even though the participants are imperfect.

Once you see it this way, Claude Code and Codex and the rest of today’s tools start to look like what they are: early tribal chiefs. They work well while the chief is competent and has a clear goal. They fall apart the moment the chief is replaced. They aren’t broken — they’re wonderful! — but they’re clearly tied to the particular engineer running them.

Some engineers will keep working this way for another decade, maybe two. But it doesn’t scale, and agents are improving too fast for that to be the long-term answer.

Three. The right words is what creates wealth.

This is the part I keep underestimating, and I think most engineers do too.

Consider the U.S. founding fathers. Without taking anything away from the millions of other people who contributed to America’s success, I think it’s fair to say that if you tried to assign credit for the country’s outcomes, the founders would hold a real slice. Not 50%. But also not 0.001%. Probably somewhere in the low single digits — call it a few percent of everything the country has produced and will produce.

Multiply that share by U.S. GDP, added up over two and a half centuries; or use standard S&P 500 annual multipliers if you wish. The number is staggering. If you want a cleaner version, use tax revenue as a proxy: assume government spending is rough evidence of value captured, give the founders a share of that, and you’ll still easily convince yourself that the founding fathers have single-handedly acted as the largest value creators in the history of humankind.

What did they actually do, mechanically? They wrote words. They sat in rooms and argued about phrasing. They put text on paper, said it out loud to the right audience at the right moment in history. And what they shipped is a document that turned an unruly population of mostly uneducated, self-interested people into a society that spread and lasted through time. That’s the whole artifact. A constitution. A set of words.

This is the skill the best engineers and architects now need to master. Not prompting our agents by hand. Not building guardrails so you can write fewer prompts. Not cutting token use. Writing the words that, said out loud to the right system at the right moment, turn a population of capable-but-aimless agents into something that creates lasting value.

Concretely, that means: every harness rule should express clean intent. That intent should cut deep — deep into what’s actually being built and why. The system around each rule should leverage all of them, ask questions when the intent is unclear, and build up understanding instead of starting fresh every time. In the background, the harness delivers everything the user asked for — plus the things they would have asked for if they’d thought of them.

Conclusion.

Two things make the combo of the three ideas above more valuable than it sounds.

First, tokens are stupidly cheap now, and they’re going to keep getting cheaper. Models will keep getting better. Hardware will keep getting better. Optimizing for token count is optimizing for the wrong thing. It’s like the founders worrying about the price of parchment.

Second, feedback loops have collapsed. In the pre-AI era, good software used to get built over decades, and even the best engineering teams measured the feedback loop from a product owner or QA in hours at best — more often days. Right now it’s hours max, oftentimes minutes, from feature request to its rollout. The window in which a vision can be passed along, understood, executed, critiqued, and redone is shorter than it has ever been in human history. The harness that captures that loop cleanly is going to be worth a lot.

So when I see someone tweaking prompts inside Claude Code to save a few hundred tokens, I don’t think “great optimization.” I think “you’re debating the font of the parchment.” The work that matters is constitutional.

And the right thing to measure it by isn’t tokens saved or prompts avoided. It’s universality: how much of your company’s intent the harness can absorb and carry forward without you in the loop on every decision. It’s writing the words, the structure, the system of checks that do that work. Saving 50% of tokens means nothing compared to building a harness that talks to your whole company nonstop, learns the shape of your organization within a few weeks, and then keeps producing the work indefinitely as a productive team member that carries your vision through time.

That’s the bet. The engineers who really get this — who stop seeing themselves as prompt-writers and start seeing themselves as founders writing the constitution of their agent tribe — are going to compound. And the work itself — putting the right words on paper for the right audience at the right time — is going to turn out to be one of the highest-leverage things a human can do in 2026.

The founding fathers would understand.