I keep meeting founders building "the X agent." The legal agent. The medical agent. The accounting agent. Each pitch is the same shape: a fine-tuned model, a domain-specific architecture, a moat made of expertise.
Then I open Claude Code on my laptop, ask it to build a contract-review pipeline over a folder of PDFs, and watch it do in fifteen minutes what most of those companies spent a year trying to specialize for.
The vertical AI thesis is a category mistake. Not because verticals don't matter — they matter more than ever — but because the part everyone is trying to verticalize is the wrong part.
Intelligence is horizontal. Distribution is vertical.
That sentence is the whole post. The rest is why I think Anthropic has effectively proven it, and what builders should do about it.
What Anthropic accidentally proved
Claude Code shipped as "an AI coding assistant." That framing undersells what actually happened. What Anthropic shipped is a general-purpose agent that happens to think in code.
Watch what people use it for in the wild and the "coding" frame collapses fast:
- pulling 50 PDFs out of a vendor portal, parsing them, and producing a comparison table
- reconciling a messy spreadsheet against a cleaner one and flagging mismatches
- writing a script that hits a half-broken internal API, retries failures, and exports the result
- analyzing a research dataset by writing throwaway pandas, then summarizing
- reading a long contract, extracting clauses, and drafting redline suggestions
None of that is "coding" in the engineering sense. It's knowledge work that happens to route through code. The same coding agent is doing legal triage, financial analysis, ops, research synthesis — across domains it was never specifically trained for.
The lesson isn't "coding agents are good at coding." It's that a coding agent is a shape of agent, and that shape generalizes to almost everything a knowledge worker does.
Human + computer = LLM + script
Here's the reduction that makes the generalization obvious.
Almost every knowledge worker today is mostly a person operating a computer. Strip the romance off the job descriptions and what's left is:
- A lawyer is a person who reads documents, queries databases, and writes documents.
- A financial analyst is a person who pulls data from spreadsheets and APIs, transforms it, and produces a deck.
- A doctor — in the administrative half of the job — reads patient records, looks up references, and writes notes.
- A PM reads tickets, queries dashboards, and writes specs.
The substrate is the same across all of them: structured data, routed through software, mediated by human judgment.
So the equation:
human + computer = LLM + script
The LLM does the judgment that used to live in the person's head. The script does the operations that used to live in the person's fingers. Put them together with a loop and a tool surface, and you've replicated the working unit of almost every white-collar profession.
This is why the same coding agent that writes Python can also do due diligence, draft a privacy policy, build a regulatory checklist, or audit a CSV. It isn't specialized for those things. It doesn't need to be. The work was always shaped like code.
There is no "medical reasoning module"
Once you've seen the reduction, the case for vertical intelligence gets thin.
There's a strong intuition that says: surely a model fine-tuned on legal text is better at law than a general one. Surely a medical foundation model beats a generic one in clinical settings.
It's mostly not true anymore. Two things have happened:
- General frontier models absorbed enough of every domain that the per-domain fine-tuning edge has collapsed for most tasks.
- The bottleneck moved from "does the model know the domain" to "can the agent do the workflow." That second question is solved by tools, scaffolding, and code execution — not by a vertical base model.
When I see a startup whose core IP is "we fine-tuned a model on radiology reports," I no longer think moat. I think depreciating asset. The next frontier model release is going to flatten that gap, for free, on someone else's R&D budget.
There is no medical reasoning module sitting inside the human brain that's separate from general reasoning. The same is true of the model. Domain expertise is information applied to general cognition. Both halves are now horizontal.
So what is a vertical agent?
The honest objection: but Harvey is winning in legal. Sierra is winning in customer service. Hippocratic is winning in healthcare. Aren't those vertical agents?
Yes — and they are the proof of the thesis, not the rebuttal.
Look at where these companies actually accumulate defensibility:
- Compliance and certification. Hippocratic's value isn't a smarter clinical model; it's the regulatory work to get an AI nurse certified to talk to patients.
- Liability and trust contracts. Harvey isn't faster than a frontier model at reading contracts. It comes with the indemnity, audit trail, and BigLaw integration that lets a partner actually use it on a deal.
- Workflow integration. Sierra wins by sitting natively inside the existing support stack — Salesforce, Zendesk, the CRM — not by training a smarter dialogue model.
- Entry point and brand. Being the company an industry buyer reflexively shortlists when their procurement team writes the RFP.
None of that is intelligence. All of it is distribution — in the broad sense: how a horizontal capability reaches a regulated, risk-averse, workflow-locked customer.
The vertical isn't in the model. It's in everything that wraps the model.
The one real exception: physical work
The thesis has an edge, and it's worth being honest about. It only holds for work that's mediated by software.
Manufacturing, surgery, construction, field service, the loading dock — these involve sensors, actuators, embodied judgment, and physical risk. A coding agent can't intubate a patient or weld a pipe. The equation "human + computer = LLM + script" doesn't apply when the human's job is to touch the world.
For those domains, vertical hardware and vertical training data are real. There is something genuinely specialized about a surgical robot or a manufacturing vision system that can't be commoditized by the next frontier release.
But for the slice of the economy that's pure information work — comfortably the majority of white-collar GDP — the thesis holds.
What this means for builders
If you're building, the practical implication is sharp.
- Don't build a vertical model. Whatever specialization you cook into a fine-tune is on a depreciation curve. The frontier labs catch up, often within a release cycle.
- Don't build a vertical agent architecture. The shape that wins is going to be the same coding-agent shape that's already winning in code. There is no special "legal reasoning loop" worth inventing.
- Do build vertical distribution. The compliance, the integrations, the certifications, the workflow surfaces, the trust scaffolding — that's where the durable value sits. It's also work the labs will not do, because it doesn't scale horizontally.
- Treat the model as a commodity input. Pipe in whatever the best frontier model is this quarter. Replace it next quarter when the price-performance shifts. Don't wed your product to a specific model — wed it to a specific customer.
This pairs with the argument I keep making elsewhere: the intelligence tide is rising on its own. Whatever cognitive gap you eke out today gets filled by the next model release. The leverage isn't in having a smarter brain. It's in being the thing a specific industry actually adopts.
The startups worth betting on right now look more like industry-specific Stripe than industry-specific OpenAI. They're not training their own brain. They're building the rails that let a horizontal brain plug into a vertical reality.
Closing thought
Every agent is a coding agent, because every knowledge job is a coding job in disguise. The question isn't what kind of intelligence does this industry need — the intelligence is already general. The question is who builds the wrapper that lets it walk through the front door.