Rebuilding the World for AI Agents — Not Humans

My work environment has changed in the last few weeks. Most of the things I used to do manually — writing emails, searching through my Drive, reorganizing my calendar — are now handled in a terminal, through parallel sessions with Claude Code or Codex.

I used to reserve this approach for software development. But a pattern has become impossible to ignore: for most comparable tasks, the machine is faster, more consistent, and often more precise than I am.

I know how imperfect this way of working is. I also notice how rarely people around me work this way. But there's a gap between what AI can technically do today and how most people actually use it — and I can't stop thinking about what that gap means.

This article is my attempt to explain it:

• Why developers are experiencing today what lawyers, consultants and doctors will experience tomorrow
• Why the real paradigm shift isn't what most people think
• What's actually at stake: the race to rebuild the world for AI agents rather than for humans
• And why that creates real entrepreneurial opportunities — for now

Almost nobody has grasped how deep this shift goes

The timeline is familiar. ChatGPT was a chatbot in December 2022. Four months later, GPT-4 was passing the bar exam. By August 2025, GPT-5 was producing mathematical proofs at a level comparable to researchers.

Model performance keeps improving. Context windows keep growing. Agents can now handle longer and longer autonomous tasks.

But the most structurally significant change isn't raw capability. It's the ability of models to interact with external tools — the web, business software, databases.

A model's usefulness now depends more on the environment it operates in than on the underlying model itself. OpenClaw made this point clearly: give an AI access to the right tools — email, WhatsApp, Chrome, Notion — and let it work. The results are surprisingly good.

The bottleneck isn't intelligence. It's the working environment.

Developers are experiencing now what lawyers, consultants and doctors will experience later

Andrej Karpathy, Simon Last, Kieran Klaassen — ex-Tesla AI director, CTO of Notion, co-founder of Cora. All three have one thing in common: they haven't written a line of code themselves in months.

Instead, they organize systems where AI agents do most of the actual development. One agent writes the technical spec. Another writes the code. Another reviews it. Another runs tests. It sounds like a management structure because it is one.

What's hard to believe is that this actually works. It works regardless of project complexity, language, or domain. The only real requirement is giving the AI the right context.

Developers can build these systems because they have a structural advantage no other profession has yet: the codebase.

A software project typically lives in a single repository, organized into files — database migrations, UI styling, business logic. Give an AI access to that folder and it understands 90% of what's being built. The context is dense and co-located.

That's exactly what isn't true for other knowledge professions.

A lawyer's context is scattered across email threads, Drive documents, case management software, in-person meetings, Legifrance, and legal research databases. An agent doesn't have access to all of that the way it has access to a repository. The context is fragmented.

This is the real reason AI produces disappointing results in law — or medicine, or consulting. Not because the model is less capable, but because it's working with a fraction of the necessary context. Fix the environment and the output changes dramatically.

The intelligence was never the problem.

The race to redesign everything for agents has already started

What's going to happen over the next several years is essentially a reconstruction of our digital world — and eventually, perhaps, our physical one — to give agents the environments they need.

Tina He describes this as a race: "The Race Is On to Redesign Everything for AI Agents." It's already underway in software.

Y Combinator recently updated its motto. "Make something people want" is becoming "Make something agents want."

The implication is significant: the value of software is no longer primarily in the user interface. It's in the ability to be accessible — not replaceable — by an agent.

That raises concrete technical questions that most people haven't started thinking about:

• Working with Word documents? The XML overhead adds layers of complexity that degrade AI reasoning quality.
• API returning JSON? Agents prefer Markdown — its structure consumes fewer tokens.
• Site on WordPress? MDX structures better than HTML if you want ChatGPT to navigate and cite you in search results.

The entire stack of digital services is going to be reinvented around agents as the primary users. This is already visible in which new products are gaining traction:

• Postgres → Supabase
• Readme → Mintlify
• SendGrid → Resend
• Stripe → MPP
• Google Docs → Proof

In each case, the question isn't "how do we add AI to an existing product." It's "what would this product look like if it were designed first for agents, with no friction?"

Why this is a real window for new entrants

Incumbents in software have built their moats partly around the UI. Users who have spent months learning a product resist switching. That inertia compounds for years.

But if agents become the primary users of software — not humans — then those carefully built interfaces become liabilities, not moats. Incumbents will have to rewrite hundreds of thousands of lines of code, every six months, as models evolve. That's a genuine structural problem.

The pricing model compounds it. Most software incumbents charge per seat. The agent economy runs on usage-based pricing. Shifting to usage-based pricing means cannibalizing your existing revenue base. Big companies are notoriously bad at doing that to themselves.

New entrants don't have these problems. They start with a clean codebase, a flexible pricing model, and no installed base to protect.

More importantly, they can build inside what I think of as the agentic flywheel:

1. Build the product with AI (development phase)
2. Let agents use the product (usage phase)
3. Get indexed in AI search engines (acquisition phase)
4. Watch the AI learn from how it uses the product (iteration phase)
5. Extract lessons from that usage (learning phase)
6. Ship new features (back to development)

Each loop is tighter than the last. The more agents use the product, the better the product serves agents, the more agents find and recommend the product.

These opportunities won't last indefinitely. Incumbents are slow, but not paralyzed. And new entrants face their own structural risk: dependency on model providers. OpenAI, Anthropic, Google — the entire stack runs on their infrastructure. That's a known tension in the economics of platform markets, and there are ways to navigate it.

I'll probably write about that next.