I first learned to give away my Legos, a concept Molly Graham from First Round coined, in 2020. Giving away your Legos is the idea that in order to scale yourself, you have to willingly hand off pieces of your job to new hires and resist the instinct to cling to what you have built.
Back then I was running my first startup and we’d grown to nearly 100 people. I had been used to knowing everything and touching every decision. That stopped working after our headcount passed 50. I had to hand off whole pieces—not just tasks, but decisions. It was uncomfortable, but it was also the only way we could scale.
Six years later, I’m learning to give away my Legos again—except this time, I’m not giving them to a team of people. I’m giving them to a team of agents.
For most of the last few years, my relationship with AI was simple: I asked, it did, I reviewed. Even when the models got good, I was still in the loop. That’s starting to change. Now the models can run for long stretches, find and use tools, and recover when something breaks. Increasingly, I give them responsibility, not prompts.
I’m not just managing models that can do things. I’m managing models that make decisions.
For example, I manage Every’s P&L, budget, and revenue. I’m responsible for matching our goals with our investments. This relies on super-accurate monthly closes (reconciling all expenses and revenue and comparing it with our forecast), and I like to do that on the first of each month. For a long time, that was manual. The most difficult part was knowing the business well enough to know where all the numbers belonged. Over months, I documented that judgment by having AI painstakingly (and slowly) edit the P&L on my behalf. Now, on the first of the month, an agent closes the books. It goes through accounts, applies those rules, and decides where items belong. I didn’t automate a repetitive task—I gave my agent the context and my tribal knowledge so that it can act like me. In short, I gave it my judgment.
If you’re willing to try, giving away your Legos to agents requires two levels of commitment:
(1) Decide what authority you’re willing to give it. Who is it allowed to communicate with? What systems can it connect to?
(2) Give it the context to make decisions the way you would. That means understanding how the business works, the exceptions to the rules, and how you make judgment calls. This isn’t accomplished in a day of work and with a single markdown file. This is months of asking your agent to do small tasks, giving it room to make mistakes, and correcting it when it does.
Giving away your Legos doesn’t happen without failure. It’s through the mistakes that you teach your agent the context, rules, connections, feedback, and permissions it needs to do it right the next time. The ones who pull ahead will be willing to train an agent the way they’d train an early employee—with context, feedback, room to act, and, most of all, trust.
