Coding Is Solved. You Solved the Wrong Problem.

Boris Cherny just declared that routine coding is solved. Sequoia agrees. So do I.

And that’s why most AI companies are already dead — they just don’t know it yet.

Here’s the death diagnosis.

The McKinsey AI survey data, reinforced across two consecutive days of intelligence signals: 94% of enterprises have active AI pilots. 7% have scaled to production. That gap has existed for two years. It predates code generation being “solved.” It predates Claude. It predates GPT-4.

If the bottleneck was execution speed, enterprise AI would have scaled by now. It hasn’t.

LangGraph published what may be the most important developer tutorial of 2026 this week — “10 Concepts Every Agent Developer Must Know.” The 10 concepts: state machines, interrupt-and-resume, human-in-the-loop checkpoints, conditional routing, memory management. None of these are code generation problems. All of them are context problems. How do you give an agent enough grounded truth to make decisions that are recoverable and trustworthy?

Nobody scales from pilot to production on code generation alone. They scale — or die — on context infrastructure.

Two categories of companies are going to fail in the next 24 months.

Category 1: Companies whose AI moat was “we ship code faster.” That moat is now worth £0. Anyone with access to Claude, GPT-4o, or Gemini has the same acceleration. If your differentiation was execution speed, welcome to commodity.

Category 2: Companies that automated the execution layer and ignored the context layer. They have AI pilots that work in demos and fail in production — because they lack the proprietary ground truth that makes outputs trustworthy. No knowledge graph. No structured domain data. No judgment layer. Just faster generation of unreliable answers.

Harvey raised roughly £160M at an £8.7B valuation this week. Harvey writes legal documents. But Harvey isn’t worth £8.7B because it produces legal text faster. It’s worth that because it understands what a specific indemnity clause means for this client, in this jurisdiction, against this counterparty. That contextual judgment is the moat. The text generation is the commodity.

Karpathy put it best at Sequoia Ascent 2026: “You can outsource your thinking, but you can’t outsource your understanding.”

The constraint for knowledge work is no longer whether machines can produce the output. It’s whether the output is grounded in enough context to be trusted.

This is what we’ve been building since February — an 86,000+ entity knowledge graph as the context infrastructure before running a single line of analysis. Not because we couldn’t automate — because automation without ground truth is expensive noise.

The uncomfortable question for your AI strategy: if a competitor had your codebase tomorrow, would your results change? If yes — your context layer is your real work. Start there.

More on this at getmanthan.com.