The bitter lesson, applied to capitalism

In 2019 Rich Sutton wrote a short essay called The Bitter Lesson. He argued that seventy years of AI research teach one thing. General methods that scale with compute beat clever, hand-built human knowledge. Researchers kept writing their expertise into systems for chess, Go, speech and vision. Each time it helped for a while, and each time search and learning at scale overtook it.

I think the same lesson applies to companies.

Companies are hand-coded heuristics

A company is a bundle of human rules of thumb about how to find customers, how to price, whom to hire and what to build next. Founders and managers write this knowledge into the system. The Bitter Lesson predicts that the bundle eventually loses to a system that finds its own rules, given enough compute and the right environment.

So the long-run winners won't be the companies with the best playbook. They'll be the companies that build the best environment for AI to find the playbook and keep improving it toward economic value.

The environment is the feedback the agents get and the definition of winning. The definition of winning matters most, because a swarm of agents becomes whatever you reward it for. If you reward revenue alone, the agents find the cheapest path to revenue, whether or not that path creates anything worth having. Profit is the signal and value is the goal. The real work is designing an environment where agents can't earn the signal without delivering the value.

The labs go direct

Today the big AI labs build these environments better than anyone. A reinforcement learning environment with verifiable rewards is a place where AI learns by trying, failing and trying again.

For now the labs sell intelligence, and founders turn it into businesses. We're the middlemen between AI and economic value, and the Bitter Lesson says middlemen get automated. The labs already sell finished products as well as models, starting with coding. Next come the markets where the most value sits. Anthropic's push into life sciences is an early sign. Within a decade we may buy medicine from an AI lab, and robots once AI moves into the physical world.

That shapes how I think about my own company. If an agent company's only asset is its agents, it competes against lab-scale compute and loses on price. The agent companies that last will combine agent speed with assets compute can't copy, such as industry access, proprietary data, licenses and customer relationships. That's where we're taking Khryseai.

Accelerate now, build the twin for later

In the short and medium term, every company should put AI into its existing operations as fast as it can. A faster organization pays off today, and the company learns what it needs for the next step.

In the long run that won't be enough. Transformation moves at the speed of people, through reorganizations, retraining and internal politics. Agent-native companies move at the speed of compute, and they get faster every year. Eventually they overtake any organization built around humans.

So the long-term move is the one Clayton Christensen described decades ago. Build the business that would kill yours before a competitor does. Next to the existing company, build an agent-native twin with as few humans as possible, and let it compete with the original.

The twin has to move at agent speed, so it can't sit inside the old organization's approval chains. It needs its own budget and the authority to make its own decisions. Few companies can build that themselves. But every established company has assets an agent company can't copy quickly, such as distribution, licenses, regulatory approval, physical assets and customer trust. The strongest twins will combine those assets with people who know how to build agent companies.

How it will probably unfold

1. The tool phase, now to about 2027. Agents are tools, and humans run every company. Small AI-native teams start beating much larger incumbents in digital markets. Few people get laid off. Companies stop hiring, and entry-level jobs disappear first.
2. The first agent companies, 2027 to 2029. Companies run almost entirely by agents appear in software, support, marketing, e-commerce and content. A few incumbents build their twin. The others keep optimizing what they have and fall behind.
3. The labs go direct, 2029 to 2032. The labs run their own agent companies wherever the work is digital. Regulated markets follow faster than people expect. Medicine is a likely early case, because AI speeds up drug validation as well as discovery, and patients will demand a drug that works against a disease that kills millions. Margins collapse for every company that has nothing agents can't copy, and the political backlash begins.
4. The settlement, same period. Governments, unions and companies negotiate how to keep displaced workers whole.
5. The physical world, from about 2032. Robots bring the same pattern into logistics, manufacturing and care. By then the main question is no longer jobs. It is that a few labs own a large share of the economy.

No single day marks the switch. The shift happens through jobs that end and are never refilled. That makes it easy to miss, and a company that waits until the shift is obvious has waited too long.

Keep paying people

If the twin works, the company should keep paying the people it replaced. The main reason is that it's fair. These people built the company, and the gains from replacing them shouldn't go only to whoever owns the agents. It also makes economic sense. Agents will make more and more buying decisions, but for consumer products the money they spend still comes from human income. Paying displaced workers keeps the economy working for everyone, not only for the owners of the machines.

Paying the full salary would be free money, and it would remove the company's reason to switch. Around 60% of previous pay is closer. In an agentic company humans mostly slow the work down, so they would keep only a small part of it, such as oversight, sign-offs and the occasional judgment call. Employees keep most of their income and get their time back. The company still pays far less than before, so it still has a reason to build the twin.

One company paying on its own can't hold up demand for the whole economy. So this will probably end up as a levy that every company pays. Germany already has a model for it. Kurzarbeit, the state-subsidized short-time work scheme that kept people employed through 2008 and 2020, pays about 60% of lost net income. The agent version would make that permanent and pay for it through the companies that replace the work.

The payments need one safeguard. A salary lasts only as long as the company does, and once the labs go direct, many companies will fail. A fund should back the payments so they keep coming if the company goes under.

Companies that start early will set the terms, and regulation will force the rest to follow.

The bitter part

The bitter part of Sutton's lesson was that in the end, the researchers' expertise didn't matter. The same holds for companies. What lasts is the environment we build, the assets we combine it with, what we reward, and whether the people who built the old economy still have a place in the new one.