Pod Insights Y Combinator Podcast The New Way To Build A Startup About this episode In the AI era, startups aren't winning by hiring faster — they're winning by automating as many internal functions as possible. In this episode of Main Function, Garry breaks down how tiny teams are beating companies 20x their size by building automations into every workflow, from engineering to ops to customer support.
Transcript 0:00 Speaker 00 If you haven't tried Claude Code in the last month, it's time to give it another shot. And if you have, you know what I'm talking about. It feels like AGI is here. One of Anthropic's own engineers writes, "Claude wrote Claude cowork." Us humans meet in person to discuss foundational architecture and product decisions, but all of us devs manage anywhere between three and eight Claude instances, implementing features, fixing bugs, or researching potential solutions. Think about what that means. The team developing one of the most sophisticated AI products in the world, something many of you probably use every day, is using this AI internally to improve their product.
0:44 Speaker 00 I think this points to a fundamental shift in how startups operate. Right now, the best teams aren't automating one or two internal functions. They're automating all of them. Often, they're tiny teams able to beat huge incumbents, thanks to internal automation. Their leanness is their superpower. I've been calling these startups 20x companies.
1:14 Speaker 00 Several years ago, my friend Parker Conrad, founder of Rippling and Zenefits, coined the term "compound startup" to describe companies that build multiple integrated products in parallel rather than focusing narrowly on one thing.
1:30 Speaker 01 The theory of the compound software business is that there's this island of product-market fit that's kind of over the edge of the horizon line that's sort of harder to get to. But if you can build multiple parallel applications at once, you can get there, and it actually ends up being a much more powerful type of product-market fit that's much harder to displace at that point.
1:55 Speaker 00 The 20x company could be an evolution of Parker's idea, but applied to internal automation. Instead of just narrowly automating a few things like writing code or handling customer support, 20x companies build automations across all internal features, code, support, marketing, sales, hiring, QA, and more. This makes each of their employees orders of magnitude more powerful than they would be otherwise. It also allows them to postpone hiring additional sales and ops staff for much longer, keeping payroll down and culture from drifting. The phrase 20x company was actually coined by the founders of GigaML, which builds voice-based customer service agents
2:40 Speaker 00 for enterprise, to describe how they managed to close DoorDash as a customer going up against incumbents that were literally 20x as large.
2:51 Speaker 02 When we got DoorDash as a customer, we were approximately like four to five engineers going against players who had like 100x engineers. So we kind of like coined the term like, hey, we are a 20x company because we are able to beat these much bigger players who are like 20x us by having a better product and better numbers.
3:08 Speaker 00 Giga was able to close DoorDash and several other Fortune 500 companies as customers because of a powerful internal agent they call Atlas.
3:17 Speaker 02 So Atlas can basically do anything within the product which you want to do. So it can use browsers, it can edit the policies, it can write code, it can do anything within the product.
3:28 Speaker 00 Atlas dramatically expands the range of what each engineer can take on.
3:32 Speaker 02 So let's say before Atlas, every engineer can probably work on four to five problems at once because they are bottlenecked by all the boilerplate stuff they have to do for the customers, right? Customers have integration, they would have to probably work on that. Now with AI FTE taking care of all the boilerplate stuff, each engineer's scope is basically doubled or tripled because they don't need to work on the boilerplate code.
3:54 Speaker 00 But Atlas doesn't just accelerate Giga's engineers, it also acts as a full-time AI employee that works in tandem with a human FTE to service dozens of accounts.
4:07 Speaker 02 Right now, we have only a single human FTE within the company. As hard as it's to believe, because we have companies like DoorDash using us, we are in pilots with multiple Fortune 500s, 10-plus Fortune 500s, where each of these companies probably have volumes over like 500,000 or a million calls a day. It's only been possible because like we have Atlas and this person can primarily focus on just the customer relationships. They ask for the customers, taking customer requests and turning them into feature requests and everything.
4:36 Speaker 00 Building an AI teammate is one approach. Another is to build an AI-integrated source of truth that gives employees instant context across your entire system. Legion Health, which is building an AI-native psychiatry network, is one example of how to do this. Legion built a custom internal interface for their care operations team that lets them pull patient history, scheduling availability, insurance codes, and a lot more.
5:04 Speaker 03 What we're showing you right now is an interface that a vast majority of our care operations team uses in their day-to-day work for anything that actually has not been yet automated. And this includes everything from, as Arthur's kind of showing on his screen, digging into a particular patient or many patients' backgrounds, trying to understand where they're at in their journey, if they need a new appointment to be rescheduled, if they're having a prescription issue, if they've sent us a message that, in traditional healthcare, might have otherwise gotten lost in the sea of different communications that go back and forth between so many different people. All of that is at a fingertips' reach for every single member of our care ops.
5:48 Speaker 00 This single source of truth interface has let Legion keep its ops headcount flat even as it's dramatically scaled revenue.
5:55 Speaker 04 So we've grown 4X in the past year, but we haven't hired a single net new person. We've been able to 4X a number of patients. We're seeing thousands of patients a month. We have dozens of providers, but we have one clinical lead, we have one patient support person, and we have one billing person. And in a typical healthcare company, those are all departments. Those are call centers. Those are groups of people sitting around desks, doing a ton of things manually.
6:19 Speaker 00 A third approach is actually build custom agents for each employee depending on their workflow and preferences. Phase Shift, which is building agents to automate accounts receivable, took this approach.
6:31 Speaker 05 So Phase Shift right now is a 12-person team, and we're going up against companies that have been around since 2006 that have hundreds of employees. The key to us as a 12-person team moving so fast is we bring AI into every process that is manual and try to automate as much as possible with AI agents.
6:47 Speaker 00 One way Phase Shift does this is by literally asking its employees to document the manual tasks they do and then building custom agents for them.
6:56 Speaker 05 So what we do is essentially say, What do you spend your time doing throughout the day? And we make them document that. And then we build quick AI agents.
7:05 Speaker 00 And this culture of relentless automation has let phase shift delay hiring for entire functions.
7:12 Speaker 05 We've actually avoided hiring a design person at the company so far to date. We're about a 12% company by just leveraging magic patterns. And our engineering team uses that to build all front end designs.
7:21 Speaker 00 These approaches aren't mutually exclusive. You can build AI teammates, a unified source of truth, and custom agents for each member of your team. The companies that do this are staying lean and setting record high growth rates. This is the new way to build. And the startups that figure it out first are gonna win.
Previous OpenClaw And The Future Of Personal AI Agents Next Inside Claude Code With Its Creator Boris Cherny Page updated October 12, 2026 · transcript and summary are machine-generated