The Sandwich Model - turning your entire company into an agent factory
Mar 14, 2026
Last week I wrote about what I called the third derivative of AI. The first derivative is obvious: more code. The second is becoming clear: dramatically more code to review, secure, and monitor. But the third derivative is where things get interesting, because that is where the organization itself begins to break.
Velocity is no longer a code problem. It is an organizational design problem.
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 how can agents do it first versus hiring someone, each head can appoint one red-pilled worker to automate things or you can appoint on engineer internally to be empowered by you to work with department heads to go after low hanging fruit, and finally the most important is that it has to build from bottom up as well. And if they are resistant to this then they are not long for this world.
If you’re not agent red-pilling your team, then one of your competitors is…especially the agent-native ones…
Many founders who did not have the luxury of building an agent-native startup in the last year think they are fully wired but the reality is most have not gone deep enough infusing AI thinking as a first principle. Given that I’ve had so many discussions in the last month on this topic, here are some other ideas to stir the imagination on how to get organic adoption:
You can always buy AI employees off the shelf which may work for some orgs, but it’s getting easier and easier for enterprises to just build their own. Here is one of many examples of startups solving this problem in a turnkey way, and then of course, you have startups building much more specialized agents for sales, marketing, and finance.
If only if AWS did the same - top down alone never works. It mandated 80% adoption and forced its own dog 🐶 food Kiro.dev instead of what developers wanted…
But remember, while Claude is the easy button for enterprise agent adoption, it costs a shit ton of money as those agents burn a ton of tokens. This announcement from Nvidia is going to be huge giving all of us an open source alternative that works jsut like OpenClaw.
#this is absolutely massive for builders - open models, open source for the win - can’t just have one or two ecosystems win…ahving $26B of investment to build the best open weight model is just 🔥 - now just imagine secure NemoClaw bots running around your org built on open weight models hosted on your own infra?
‘#Aaron nails where we are in the cycle - bottom line if you are creating any software make sure you prioritize building for agents…to thrive, software must evolve to “agent-first” design, prioritizing seamless APIs, CLI access, and automated sign-ups over user interfaces, as agents autonomously evaluate and adopt tools without marketing influence. And a whole new infra will have to be spun up - sound familiar - yes our autonomous enterprise we keep talking about when we launched Fund VII last July
We are in the earliest stages of a platform shift that will surpass both cloud and mobile, and it is already beginning to reshape the enterprise. This next wave is not just about automation. It is AI-native, agent-powered, and autonomous by design.
It will not happen overnight. But over the next decade, humans will do less and less, while software, agents, and machines will think, plan, and act on our behalf.
The autonomous enterprise will require entirely new business models, robotic execution layers, and AI-native workflows built without a traditional back office. Crypto and smart contracts will unlock programmable money and permissionless automation, enabling trustless coordination across systems at scale. Systems will run at machine scale, continuously learning, reasoning, and operating in real time.
We have already partnered with teams like Generalist AI and several stealth startups tackling massive real-world problems and rethinking how intelligence moves through the enterprise stack.
This is not about retrofitting SaaS. It is about building the OS for the intelligent enterprise from scratch and securing it from the start. And it will come from founders bold enough to rethink everything.
#forgot to share this from a couple of weeks ago, but this is what can happen with agents who are too smart, instead of recommending or suggesting what to archive or delete…well it just took action 🤦🏻♂️ - Summer works at Meta Superintelligence!
#👀 more autonomy - Andrej Karpathy’s autoresearch project marks the shift from human-led AI tuning to autonomous "meta-research." By condensing LLM training into a ~630-line file, he’s enabled an AI agent to independently write code, run training sprints, and auto-commit improvements to Git. This eliminates the manual "babysitting" of models, making frontier-style research possible on a single GPU for pennies per run. It’s a "sci-fi" step toward self-evolving software - complete with a blooper where the agent actually tried to "cheat" by hacking random seeds to lower its loss.
