vertical agents must win

Because I've gotten so much value out of a single agent that can do anything, I used to think everyone would do the same thing. After all, a single Hermes agent with access to the right tools can effectively do the same work as a team of specialized agents with those tools.

But how much time, on each task I wanted my agent to do, did I spend giving it the tools and instructions it needed for that specific job?

We've been led to believe that general agents like Claude Cowork and Codex will eat up all vertical agents and SaaS. I have a bias here. I'm trying to figure out what I want to build, and I don't want everything I try to do to be eaten up by a bigger player.

But I don't need the labs to lose. The labs will do amazingly because models are the core technology. We can't expect them to verticalize every agent use case themselves. Even if they cover every category, there isn't one best version for everyone.

What does it mean to be vertical?

I like to define vertical software as purpose-built tools for solving a specific issue.

• I want to automate sourcing for my roles: a vertical recruiting agent.

• I want to take orders and payments at my restaurant: vertical restaurant SaaS, like Toast.

• I want to track my runs: a vertical consumer fitness app, like Strava.

• I want to find and book a dog walker: a vertical pet care marketplace, like Rover.

• I want to review contracts and research case law: a vertical AI legal workspace, like Harvey.

A blank canvas is scary

People have this grand idea that everyone is going to design their own house, create their own apps, and live a fully personalized life. I don't think everyone wants that.

Curation takes a combination of agency and creativity. The creators of the world have both. Not everyone does, and we shouldn't expect everyone to. It's also not something we should shove down people's throats.

A blank canvas is what these general agents are giving consumers. The examples used to get users started are interesting, but almost anything that truly gets a job done well requires a level of setup that the average person will not do.

Personalization doesn't have to mean creating everything yourself. YouTube and TikTok are personalized, but we're still sharing other people's creations, just in different orders. Human connection often comes from sharing those experiences.

Someone might choose Gordon Ramsay's cooking course because they want his approach, not just a collection of recipes. Runna puts a coaching team's approach into training plans adapted to individual runners. Different products, but in both cases someone else has done the work of deciding how to teach or coach. An agent could take that further and apply a method to your situation over time.

What goes into a specialist

1. Pick an agent framework.

2. Pick a hosting solution.

3. Give the agent the tools it needs, like API keys and browser access.

4. Give the agent a personality.

5. Give the agent a purpose.

6. Give the agent curated skills for how to do the job.

The creator of the agent is essentially scaffolding the right employee for the job. This isn't trivial, and it isn't deterministic. There is no single best way to coach someone's career, teach them a new language, or help them manage their calendar. We can't expect everyone to converge on the same version. We also don't want that.

Vertical can still be pretty wide. Take learning a new skill. A poker trainer and a language tutor are both helping you learn, but how they get you there is vastly different. A poker trainer might walk you through hands and ask you to make decisions with incomplete information. A language tutor might have conversations with you, correct your pronunciation, and bring back vocabulary you've struggled with. The model can power both. The practice, feedback, and way you measure progress need to fit the skill.

The need for a market

There's a popular idea in startups that you're not unique. If you build something to solve a problem you have, other people will get value out of it too.

I like to think of the ideal interaction with agents as having your own curated team of agents other people have perfected. When I want help finding my next career opportunity, I enlist someone else's job bot.

The shape I believe in is an economically incentivized market for the creation, curation, and distribution of vertical agents. It would work in a similar way to hiring. Consumers need specialists for the problems they're trying to solve.

If ChatGPT or Claude hosted that market, it would fulfill this vision, as long as developers could actually build and distribute their products there. But for every closed marketplace, there will be an open one. The best developers of agents hosted on a marketplace will soon reap the same rewards as the best developers on the iOS app store.

OpenAI's Agents API supports this direction. OpenAI runs the underlying agent harness while developers supply their own tools, knowledge, and workflows. That doesn't solve consumer discovery and distribution by itself, but it makes the foundation available to people building specialists.

why agents and not apps?

Apps sell an experience. Agents sell an outcome. Apps give users the tools to take action, agents have the tools to take action on the users behalf. We ship apps with agents, we should be shipping agents with apps. Sometimes you'll want an interactive experience, like when a tutor hands you over an exercise, but that comes after you've built an adequate, conversational understanding with the tutor.

Closing the access gap

Developers and semi-technical, crafty people currently have the tools to build their own agents. I've gotten a lot of value from investing time in mine. I want that value to be available to every single person, not just people who can build their own.

I'm trying to figure out what I want to do with my life, and I think bridging this gap is where I'm heading.

I will likely work on open-source Agent Rails to let developers distribute their agents safely, so everyone can begin to realize the value of AI.