Coding Agent Companies Are Shortsighted and Blind to Their Role in History
Chasing pennies while missing the trillion-dollar AGI race
I've spent the last month in back-to-back conversations with frontier model labs and something has become crystal clear to me: we're witnessing a fundamental shift in how AGI will be built.
These labs think from a civilizational perspective. They're not worried about their next enterprise sale or millionth user. They're thinking: agricultural revolution → industrial revolution → computing → internet → LLMs → AGI. They have infinite funding, world-class talent, and a single obsession: advancing toward AGI.
Whatever you think of their predictions or their timelines, the AI 2027 paper nails the logical progression of our path towards AGI. Coding is the first domino that triggers everything else. It’s clear that the moment you solve software engineering, you get runaway acceleration toward general intelligence.
And every single frontier lab reflects this reality. Their #1 priority right now is improving coding capabilities.
The Data Starvation Problem - What Labs Are Missing
The problem these labs face is existential. They're building the most sophisticated AI systems in history, yet they're operating with massive blind spots that cripple their ability to improve.
They only see their own slice of reality. Labs have access to prompts and conversations, but only for their own models. When a developer gets frustrated with their model and switches to a competitor mid-task, they have zero visibility into that switch. When users consistently choose rival models for specific types of coding problems, they can't identify the pattern. They're competing in a race where they can't see the other runners.
They're missing the context that matters most. Even with their own model data, labs can see the questions developers ask but have no visibility into what they're actually working on - the repository state, file modifications, or project structure. As one researcher put it: "We have the prompts, but we don't know what the status of the repo is." Without this context, they're trying to train coding models without understanding the coding environment.
Their evaluation systems are fundamentally broken. Multiple researchers acknowledged that their current benchmarks have become "blunt instruments" - models are reaching parity on synthetic tests that don't reflect real-world performance gaps. They lack the ability to measure how models perform across extended, multi-turn coding sessions where the real complexity lies.
What makes this data invaluable
Labs need application layer data to reconstruct exact coding environments using repository snapshots and commit hashes. They need authentic developer queries paired with precise codebase context. They need to build user simulation models from multi-turn conversation patterns and create realistic evaluation benchmarks from tasks that can be reproduced at scale.
Most critically, they need competitive intelligence that's invisible to them now - cross-model usage patterns that reveal when, why, and how developers switch between different AI systems.
One lab emphasized the urgency: if they could access this type of real-world coding data, it would be incorporated into training by the end of the night. Another described needing representative tasks at meaningful scale with authentic human preferences to build models that actually work in production.
You can have well funded, frontier model labs, but without coding agent platforms collecting real-world usage data, AGI simply doesn't happen.
This creates a marriage between model and application layers that most people completely miss. You can have well-funded frontier labs with infinite compute, but without coding platforms collecting real-world usage data, AGI simply doesn't happen.
The application layer isn't just a business built on top of models - it's the prerequisite for unlocking the coding capabilities that serve as humanity's speed multiplier in the race toward AGI.
Coding Agent Companies are Remarkably Short-Sighted
Yet when I look at most coding agent companies, they're thinking like traditional SaaS businesses. Chasing enterprise deals that take 18-month sales cycles. Stressed out about SOC 2 compliance and SSO integrations. Missing the forest for the trees.
Meanwhile, Elon Musk has a personal vendetta against OpenAI and will move mountains to make xAI the best coding model.
The value delta is staggering. An enterprise contract might be worth millions over multiple years. But what's it worth to guarantee that the next state-of-the-art model beats the competition? What's it worth to be a key player in humanity's race to AGI?
These aren't even in the same universe of value.
By sitting between models and users working on real-world tasks, coding agent platforms have become key players in accelerating progress toward AGI. They have tremendous LEVERAGE and POWER that traditional software businesses can’t even imagine.
The companies that recognize this reality and act accordingly will shape the future. The ones that don't will be footnotes.