What's missing on purpose

(This is a story about Agent Fang, a live app — you can open it here.)

The first version of Agent Fang could do far more than the one online today. It was a real agent — it could hold a conversation, pick its own tools, switch between Claude and GPT, and yes, count the r's in "strawberry," the little task the famous chatbots are mocked for getting wrong. It had everything serious software is supposed to have: a database, a memory of past conversations, monitoring, the full cloud setup. I was proud of it. Then I deleted most of it.

Last time I wrote about software that works and still goes unused — built, demoed, quietly abandoned. Cutting my own agent in half was how I tried not to build another one of those.

What is online now is a single flow. No memory of past sessions, no dashboard of metrics, none of the machinery that lets software scale to thousands of users. I took all of it out, and not because it was bad work. I took it out because none of it answered the only question a new user actually asks: can I see what this thing is doing? Every piece I removed was a piece standing between the user and that answer.

One thing I kept, and would never cut: the front door. The app is open on the internet, and anything open on the internet gets abused — so it sits behind the same kind of security gate a bank's website uses. That's the wall subtraction runs into. You strip away what hides the work from the user; you never strip away what keeps it safe.

What's left reads like a recipe: intake, validate, rewrite, search, select, fetch, read, synthesize, save, email. You can follow it top to bottom. It is not a brain making decisions you cannot inspect; it is a sequence you can point at. On screen, each step is a chip that lights up as the work moves through it. When something breaks — and it does — you can see which step, and why. In a business that visibility is not a luxury. It is the line between a tool people trust and a tool they stop opening.

If you are weighing an AI tool for your own shop, this is the part worth keeping: you do not need the heavy machinery to get the value. The scaffolding that lets a system serve ten thousand users is the same scaffolding that hides it from the first ten. Take it away, and the part that does the real work stands in the open.

That openness cuts two ways. Because the flow is just a readable file — the steps, the order they run in, even the small piece of code behind any one of them — it is fixed while it runs but cheap to change between runs. I can add a step or reroute the whole thing without unpicking a model's reasoning. So can someone who has never seen the code before. That turns out to matter more than I expected.

None of this comes free. A flow this plain only works when the job is well-defined — when you can name the steps ahead of time. Ask it something open-ended and a straight recipe has nowhere to go. But the fix is not simply to turn loose more autonomous agents. More often it is to let the flow itself handle a little more — a fork in the road here, a step that tries again until the result is good enough — while keeping the agents few and on a short leash. That balance is the whole job: enough in one flow to do the work, not so much that you can no longer follow it; enough agents to cover the gaps, not so many that you cannot tell what each one is doing. Too little and it cannot handle the task; too much and you are back to the opaque machine I just deleted. The balance is a judgment about your business, not the technology — and no vendor will make it for you.

So here is my question this week: which task in your day is actually narrow enough to become a flow you could watch run? Reply with one, even a rough one.

And it leaves a question I am not going to answer today. If the flow is just an editable file, who keeps rewriting it as the work changes? These days, mostly not me, and mostly not by hand. That is where this goes next.

Lollapalooza is coming up here in Chicago, and the kids are scrambling for tickets. Enjoy the summer.