How Ramp turned a 1,200-person company into an agent factory
Apr 11, 2026
The discussions on how to go wall-to-wall with agent building and deployment across every organization continues every week. I’ve already written about this a couple of times in the last month from What’s 🔥 #488 (When Product Velocity Breaks the Company) and #489
This was the dominant topic at board meetings this past week. And the answer, IMO, is the sandwich model. Agent red-pilling your company has to come from both ends. The founder/CEO drives from the top, but real ownership only comes from the bottom up, through organic usage. The question is how.
Every company does it differently. In the sandwich model, it has to start with the CEO, then every department head needs to be bought in 100% in thinking ...
Every board meeting right now has the same conversation: coding first, yeah, but then what? How do you agent red-pill the whole org? And the question I keep getting is: show me someone who’s actually done it.
Here’s the best example I’ve seen yet. Ramp. A company founded in 2019 with over 1,200 employees is more agent-pilled than most startups founded in the last 12 months.
The numbers alone are staggering: AI usage up 6,300% year-over-year, 99.5% of the team active, 84% using coding agents weekly, 1,500+ apps shipped in six weeks from 800+ different builders. Non-engineers now account for 12% of all human-initiated PRs on the production codebase. This is the Autonomous Enterprise in action.
Ramp didn’t just adopt AI.
They rebuilt the company so that everyone can ship software.
But the numbers aren’t what’s interesting.
What’s interesting is how.
It validates everything I’ve been writing about in the sandwich model. Agent adoption has to come from both ends. The CEO drives from the top, but real ownership only comes from the bottom up through organic usage. Ramp did both. Leadership set the expectation. But they also built the infrastructure for people to teach themselves and each other.
The key insight that most companies miss: the harness matters as much as the model. They hit 99% AI adoption and then noticed something alarming. Most people were still stuck. Not because the models weren’t good enough. Because terminal windows, npm installs, and MCP configurations were too much for most people, and the few who pushed through had siloed setups with no way to share what they’d learned.
So they built Glass, their own Claude-powered agent workspace. One SSO login, 30+ tools pre-connected, zero setup. 700 daily active users within a month of launch.
The biggest lesson: the people who got the most value weren’t the ones who attended training sessions. They were the ones who installed a skill on day one and immediately got a result. Get people to the “aha” moment as fast as possible. The product teaches faster than you ever could.
The other thing they got right was org design. My sandwich model talks about top-down mandate meeting bottom-up organic adoption. Ramp’s version: a small central team builds the platforms and plumbing, functional teams build on top and give feedback that drives the central roadmap. The spokes drove the center as much as the center drove the spokes.
In a traditional SaaS company, software is built by engineers and used by everyone else.
In an AI native company, everyone builds.
Sales ops builds tools. Finance builds workflows. Support builds automations. Engineers still build the platform, but the rest of the organization builds on top of it.
When that happens the entire company becomes a software factory.
And once that flywheel starts spinning, execution speed compounds in ways that are very hard for competitors to catch.
The biggest surprise from Ramp? It wasn’t who built the most. It was how many people had been waiting for permission to build at all.
That’s the real unlock. Most employees have far more capability than their companies give them credit for. You have to be intentional, you have to make it easy for everyone, and you have to show the way.
It’s also no surprise that Ramp is consistently named one of the best companies for future founders to work at. They hire and screen for ex-founders or potential next founders. The builder culture isn’t an accident. It’s a hiring strategy.
Which brings me to DoorDash. If you're a SaaS company watching all of this and still debating your AI strategy, you'd better pull a DoorDash.
They just acquired Metis, a ~10-person YC S25 company that hadn't raised institutional capital, for $150M all cash. Why? To go all in on agents and bring that talent in-house. Hard to find the best agent builders when they all start their own companies, so sometimes you have to buy them. Pay up like your life depended on it, because in this environment, it might.
And this isn't just an operating story, it's an investment story too. Nearly half of all private equity deals now target software and technology services companies, a share that has doubled over the last 15 years. The pressure on those portfolios to agent-pill their orgs and defend their value is enormous.
#the Mythos moment was last week with the announcement of Project Glasswing - A model too powerful to release because it can find vulnerabilities better than almost every human on earth which is simply amazing and terrifying.
How powerful? In one case, Mythos found a 16-year-old flaw in widely used video software, in a line of code that automated testing tools had executed 5 million times without catching it. The Treasury Secretary and Fed Chair summoned major US banks to discuss the cyber risks. Bank of Canada did the same. This isn’t a tech story anymore. It’s a national security moment.
#but Wall Street doesn’t get it - no large enterprise is automatically going to just let frontier models and agents run wild patching everything with understanding context and dependencies, formal verification and audit trails, and some human oversight - for the simple fixes, yeah, but the harder ones, no way! Hence, a massive overreaction on some of these stocks
#i look forward to this every year - Jamie Dimon’s annual letter which covers everything from markets to geopolitics to our economy and finally tech - here’s the AI angle
#10 employees bought for $150M - Metis is an applied-research and product lab building proprietary intelligence: the post-training and continual-learning layer for enterprise agents.
#😱 every data source an AI agent touches is an attack vector. Every one. Google DeepMind just tested 23 attack types across frontier models and the defenses we have today fail. This is the next massive security category waiting to be built
#🎯 Aaron nails it - no model can safely do “continual learning” across an enterprise. One banker’s docs are invisible to the next. Sanitizing secrets is impossible. The context layer isn’t optional - it’s where general AI actually becomes useful.
