Most companies treat AI like a chatbot. You open a window, you ask, you copy the answer, you close the window. Block did the opposite, and that is why what they did is worth looking at.

Block built an internal system called Builderbot that coordinates AI agents across its entire codebase. It is not a tool an engineer opens. It is something that lives inside Slack, where its teams already discuss the work. And that is the thesis of this post: the interface that wins is not a chatbot or a new IDE, it is the conversation where the work already happens, and the value is not in the model you use but in the system you build around it. The second part does not come prebuilt. You build it, and that is where we come in.

What Block put to work

Start with the concrete uses, because that is where it clicks fast. An engineer tags Builderbot in a Slack thread. The system researches the problem, builds a plan, and ships the change. It is not an assistant that suggests; it is a system that executes end to end inside the flow that already existed.

The numbers Block published give the scale: 200,000 operations a day, 1,500 PRs merged a week, and 15% of all production code changes at the company. What used to take months now takes days, they say.

Jack Dorsey framed it by saying they are going to talk a lot more about their “intelligence tools,” and that this is just the beginning. Strip out the founder enthusiasm and something concrete remains: a company that stopped treating AI as a feature you buy and started treating it as a system you build.

The interface that wins is the conversation you already have

Here is the part almost nobody copies well. The question of “where do I put the agent?” almost always gets answered wrong: in a new app, in a separate panel, in another tab the team has to remember to open.

Block put it in Slack, and that is not a detail. As Sergey Karayev noted commenting on the launch, the best context they found for coding agents is nothing sophisticated: it is the Slack thread. The problem is already explained there. The screenshots, the debate, the prior decisions, even the “we already tried this and it did not work”, are all in it. An agent tagged in that thread receives all that context without anyone rewriting it.

A new app does the opposite. It forces the user to recreate context that already existed somewhere else. And that step, rebuilding context by hand, is exactly where most AI pilots die. Not because the model is bad, but because you asked it to work blind. I argued this at the time from the angle that AI is not a tool, it is infrastructure: infrastructure is not something you visit, it sits underneath everything you already do.

The number that matters is not 1,500 PRs

It is worth pausing on a criticism that showed up the moment Block published the figures: that counting PRs is an empty metric. They are half right, and the half they are right about matters.

PR count on its own can be noise. Fifteen hundred changes a week that never reach production, or that create more review work than they save, are not a win, they are a problem dressed up as productivity. If all you measure is volume, you are measuring the easy thing, not the thing that counts.

The number that does matter in Block’s case is a different one: 15% of production code. That does not measure how much the system generates, it measures how much they dare to let through. It is a trust signal, and trust is the hard thing to build. I covered this same tension when Uber reported that 70% of its code is already AI: the figure impresses, but what decides whether it helps is the review and orchestration system behind it, not the percentage on the slide.

What IQ Source does about it

The lesson for your company is not “install an agent in Slack.” It is understanding what kind of thing you are building. A chatbot you buy and connect. A system you build, because the part that creates value (the tools you give the agent, the context it receives, the limits on what it can do alone) is specific to your operation and does not come in a box.

That is what we stand up when a company puts us to work building on top of their operation. We do not hand over a chatbot license. We build the orchestration layer on top of the model, connected to the tools and processes you already use, placed where the work already happens. The model is swappable. The system around it is yours, and it is what keeps creating value when the fashionable model changes next quarter. I wrote about that in the harness is the moat: the model is a commodity, what you build around it is not.

Before you buy the next AI license for your team, ask a different question. Not “what tool do I buy?” but “what system do I build, and where do I put it so it gets the context I already have?” The first question leaves you with one more tab nobody opens. The second leaves you with something that becomes part of how your company works.

Let’s build the system, not another chat window