Want a billion-dollar agent startup? Start with your own job

Welcome back to Building AI Agents, your biweekly guide to everything new in the field of agentic AI!

Or, for those who want to avoid kidney failure, just take a twice-weekly dose of Building AI Agents.

In today’s issue…

• Build agents from a description with Google’s new tool

• An open-source alternative to the most powerful thinking agents

• How Zapier’s founder saved 20+ hours a week with agents

• Will AI agents eat the SaaS market?

• Price your agent like you mean it

…and more

🔍 SPOTLIGHT

I wanted to do something a little different with today’s spotlight.

Normally, the inspiration for these is some news item about agentic AI or a general thought about the trajectory of the industry, but this week it’s this post by Y Combinator partner Jared Friedman.

Last week, I wrote about some of the agent startups that have hit multi-hundred-million or billion-dollar valuations in a short period of time, and offered some advice on how one could achieve a similar success story: pick a language-heavy (since these are large language models we’re talking about) task that is a major pain point for lots of professionals and businesses, and automate it with an agent.

I definitely stand by this, but I want to add a corollary: the best of these pain points to automate is one from your own job.

On the surface, the reason for this might seem pretty obvious. If you do something every day, it’ll be easy for you to build an agent that does it, right? Partly true, but it’s actually not the main reason. Believe it or not, building an AI agent is often the easy part, especially with so many code and low-code frameworks out there, plus a host of vibe coding tools to create agents from scratch. Give me the outline of any task that mostly involves doing stuff with text, and I can build an agent that can take a pass at it within an hour or so.

But getting that agent to work well is a different story. If you show the output of one of these built-in-an-hour agents to a domain expert whose work I’m trying to automate, they’ll immediately recognize all of the flaws in its output, flaws that would totally escape me. This is the actual hardest stage of agent building: evaluating its output and iteratively adjusting its prompt and architecture until the work being done is satisfactory. Only someone who actually does that work for a living can truly judge what that looks like.

I suspect this is why so many startups have sprung up to automate software development. Who starts software startups? Software engineers. And what job do software engineers best understand? Software engineering. Clearly, this approach worked for the Lovable folks I wrote last week’s piece about, whose vibe coding agent startup is now doing $100 million in annual revenue. The world needs a lot of software engineering.

Which leads me to the final, more fundamental reason why automating your own work is the best place to start. If you’re building an agent to do parts of your job for you, those parts are probably your biggest pain points, the things that suck up the most of your time and give you the greatest aggravation. Unless you have a very uncommon job like pet obituary writer, there’s a good chance that millions of other professionals like you have the same pain points. I started this by saying that automating a major pain point for lots of professionals and businesses is the path to a hyper-successful startup. You might pay $50 a month or more for an agent that saves you an hour a week. Multiply this by the million or more people who have the same job as you and…you see where I’m going with this.

So if you want to build AI agents, start with what you do every day. You might save yourself a little bit of time and annoyance. Or you might end up with the next billion-dollar agent startup.