Code Has Become Cheap: AI Review Is the New Bottleneck

I haven't posted lately. I've been very busy. I've been publishing posts like this in blog form on my portfolio site.

So, to get to the topic:

Rice-Harvesting Machine — Far Too Much Waste

If writing code itself used to be a major bottleneck, that is no longer the case today. We used to harvest rice by hand and separate it from the chaff by hand, but now that rice-harvesting machines have become incredibly widespread, they produce far too much waste.

The same is true of code. Producing it has become very easy, and reviewing the code and separating the wheat from the chaff is now the main bottleneck.

This pattern was articulated more clearly by Andrej Karpathy, one of the founders of Tesla and OpenAI, at the 2026 Sequoia AI Ascent. Speaking about himself, he noted that until November 2025, the share of code he wrote by hand was ~80%, but starting in December, that reversed, with 80% being delegated to agents. He stopped calling it "vibe coding" and gave it a new name — agentic engineering. The key difference is: "You don't write the code directly; instead, you direct agents and provide oversight."

In other words, we have managed to make code production cheap. What remains is sorting it.

Where is the world headed?

Many companies hold weekly meetings. That was where plans were made and presentations were given. Why? Because writing code used to be expensive. Before making a decision, you first had to talk, draw diagrams, and reduce risk.

Today, that picture is changing. More companies are now holding meetings by showing a working prototype or MVP. Vague ideas and plans are now worth $0.

Build it and bring it in. If it doesn't work, just make it work. If it was built wrong, just fix it and rebuild it — we have reached that point.

The new default is that you are not writing the code directly 99% of the time, you are orchestrating agents who do.

AI Native Organization — a new holistic culture

This culture is itself a set of rules and policies that the best YC Startups and top technology companies have begun adopting. We call this an "AI Native Organization". Not only in Mongolia, but even most companies worldwide have yet to adapt.

The defining characteristics of this culture:

• Sells services, not software. An AI-native company will not give you a tool — it will do the work for you. Because the global "services" market is many times larger than software, the main opportunity lies here.
• Capped by compute, not headcount. One YC W26 graduate, Light Anchor, builds e-commerce brands run entirely by AI agents. Its operations are not constrained by headcount. Its internal operations are all handled end-to-end by agents.
• "Company brain". A new primitive recently defined by YC: all of a company's knowledge (documents, Slack, plans, design, codebase) should be brought together in one place as a structured, self-updating "AI execution skills file". Otherwise, there is nothing agents can do.
• Fewer interns, more output. CharacterQuilt can now build, design, and deploy a marketing campaign in 1 hour—a campaign that previously required 3 agencies, 10 tools, and 6 weeks.

The internal standard at these companies is always the same: "Build it and bring it in before talking about it."

Enterprise systems are not that easy

Unified user management around the core product, permission settings, audit log, custom domain, billing, SSO... and so on. Developing these can sometimes require more development work than the core product itself.

To put a number on it: according to research published by WorkOS, building enterprise-grade SSO + multi-tenancy in-house requires an average of 12-16 weeks of dedicated engineering — and that is only the "login works" part. On top of that:

• Multiple identity providers (Okta, Azure AD, Google, OIDC, SAML).
• Tenant-level config: custom domain, branding, attribute mapping.
• Noisy neighbor problem: when one tenant starts a heavy workload, everyone else's performance drops.
• Internal admin queries that "scan all data" crop up.
• Cache, queue, job — all of them need to understand tenant_id.

This is not a "feature", but infrastructure. If you build this infrastructure incorrectly, you may end up throwing away the code for the actual product and starting over.

Personally, I am now actively researching how to implement this system. I will continue sharing my findings going forward.

The next Microsoft

The best time to build the next Microsoft is "now". Deadline: 1 year. In 1 year, the market will probably already be saturated.

Why 1 year? Because:

• The agentic engineering tool chain (Claude Code, Cursor, replit agent, OpenAI Codex CLI, etc.) is becoming GA in 2025-2026. The people who will use it are ready.
• The YC W26, S26 batches are entirely AI-native focused. Top talent is pivoting as pairs.
• Enterprise "AI transformation" budgets are becoming actively available (the emergence of "AI-native consulting" companies like Foaster from YC is proof of this).

So, let the race for the market begin. 🐎

---

Sources that helped with writing parts of this post:

• Karpathy on agentic engineering & Software 3.0 — Sequoia AI Ascent 2026.
• YC W26 / S26 startup directory (Light Anchor, Hedge, CharacterQuilt, Wato, Foaster).
• WorkOS — "The developer's guide to SaaS multi-tenant architecture".
• Scalekit — "Build vs. buy: SSO strategies for B2B SaaS".

In my next post, I will explore the topic of "How to build an AI Native Organization in Mongolia" in greater depth.