Skip to content

The Two-Week Bet

Posted on:Nov 13, 2026

Yesterday at 11am I spoke to a principal software engineer at a company that has been around for a while. Very good conversation. One thing was obvious from the first ten minutes: she and part of her team are still careful about AI.

I get it. People who like going under the hood and want to understand how things really work still see a lot of value in that, and I praise it. The people who know how things work should still be the ones building.

That same week I had spoken to a VP of Engineering at Goldman Sachs and a senior member of technical staff at Bloomberg. Completely different conversation.

We talked about how we do agentic engineering in the team I lead. We simplified our infrastructure and own most of it now, which has taken our costs to almost zero and made us more ambitious, because with AI you can go deeper into the stack than you planned to.

You can take an open source tool that replaces something like Ably, put it on a machine you own and test it in ways nobody thought to. You can write scripts to push it to its limits, implement algorithms from papers that came out last month, and get it doing things it couldn’t do when it was written.

Moving off vendor-locked systems like AWS onto machines you own costs about 10x less for the same workloads. The bigger win is that you’ve tested every part of it end to end, so you know exactly how it works and you can build workflows that fit your system.

For example, we swapped GitHub’s hosted runners for Blacksmith’s build mechanism in an afternoon. When our builds get bigger we’ll swap it out again, and that will take an afternoon too.

That’s the point. Be ambitious at every stage, for the problem you’re solving at that stage.

The Bloomberg engineer told me he has moved a number of things onto machines he owns, with agents securely holding the API keys to his Cloudflare and Hetzner accounts and setting things up for him. He still uses his own judgment. The agents assume a lot, so you spend a few minutes reading the plan and steering it somewhere better.

The old friction is gone too. Before, it was “I don’t know that tool, I can’t pull this Docker image because I’m not sure how it works.” Now the agent reads the README in seconds, tries three tools at once and picks the best one for the job.

So that’s the technical conversation, and getting buy-in from people who understand the stack is the easy part. It’s the same point I made in dumb agents are very useful: you don’t need the smartest model, you need the right workflow around it.

The harder conversation

Non-technical founders are the harder conversation. They have deep industry experience, and you’re the enabler who tells them what the technology can do. They’re standing at the edge of the water, not sure AI can carry the weight.

What works for me is making the bet small.

Tell them we can prove this in two weeks and you’ll show them the results. If two weeks isn’t enough, we go back to the boring traditional plan we were going to do anyway. They’re only risking two weeks.

An hourglass testing the water with one toe: two weeks, not two years

If you win those two weeks, the next six months is a labour of love. They’re sold.

Then something else happens. Once non-technical people are sold and see the vision, they start to over-ask for what the models can do. Now you’re balancing the other way, and you have to be clear, early, about what the models can and can’t do.

I’ve seen this across multiple teams now. It’s very early, and most non-technical people haven’t seen what you can really hand to AI in a real workflow. The moment they see it, it clicks and they want more.

A chat window lifted off its feet by a balloon: ambition runs ahead of the models

You might be thinking: but I used Lovable the other day, and I built many cool and useful things. Yes, but the problems these teams deal with are of a different shape. They’re problems that require you to collapse all the screens in your Lovable app into a chat with durable memory and proactive, always-live behaviour, or apps that take 90% of your work and get it done for you.

For example, imagine an Airbnb management product that needs to handle paperwork for you on old, clunky websites, and has to manage communication between landlords, property managers, Airbnb itself and users, with only 2 screens for the landlord to check progress. It has to learn from the general data available to it and figure out the white space to solve for. This one I will talk about in my next issue of this newsletter.

It’s a different type of hard problem.

I’m not 100% sure of the whole solution yet, but the first step is the smallest prototype that can win the two weeks. Set the goal and get it sorted.