Talk is cheap.

For six years, my job was getting people to click buttons. I built a company around it. I learned how to design a tour that got finished, a checklist people followed, a tooltip you wouldn’t automatically dismiss. I was good at it. I thought that was the job to be done.

Then AI ate software, and I went to build the other side: agents that do the clicking for people. We’re not teaching users where to click anymore. We’re teaching agents how to use software.

I expected a hard project. It turned out to be a different kind of hard. Getting an agent to reliably operate real software, read the right state, take the right action rather than the one beside it, know when to stop and ask a person, and behave on run five hundred just as it did on run one is not a feature you can vibecode. Every time I thought I could see the bottom, we found another turtle below it. As it turns out, gettig the chatbot to chat was the easy part.

And here is the part that actually scares me. Agents are already reaching into products and operating them whether the companies behind those products set the rules... or not. I have watched one move through a settings page, slowly bumbling along and leaving no trace anyone would think to inspect. That is the today version. The tomorrow version is slicker and worse: when the agent does the work from the saftey of codex or claude, the human stops coming to your product at all. Your relationship moves to wherever the agent lives, and everything you built becomes plumbing behind someone else’s assistant.

A support leader who has run large service queues put it to me plainly: too often, the AI is bought by the company but borne by the person who has to use it. She said we have quietly rebuilt the old phone tree, the press-one-press-two wall that exists to wear you down until you give up and go away, except now it wears a friendlier face and calls itself resolution. She was not describing a far-off risk. She was describing tools her own company is being sold today. When the people closest to the customer are already this clear-eyed about what is being done in their name, you stop being able to tell yourself the market will sort it out.

So here is what I believe. Talk is cheap. The job was never to answer the question; it was to solve the problem! And for a long while, answering accurately seemed the best software could do. But, technolology moves on. Your users do not want to learn the quirks of your product works. They are not going to read the manul. They won't click on the tour (sorry). They want the outcome, delivered by something acting on their behalf that shows its work. The layer is getting built. The question is who governs it: the company that made the product, or whatever agent showed up first.

I think it has to be the company. And inside these companies, the old org chart has stopped predicting who does the work. One of the most interesting people I met was an engineer inside a go-to-market team, a former support rep who taught himself to build with AI rather than watch the work happen without him. He is technical and close to the customer. Until recently, those were two people in two departments. That wall is coming down. Building software got cheap almost overnight. Knowing what to build, for whom, and how to do it without breaking something did not. These workflows do not belong only to engineering, and they do not belong to a model vendor that will never know your edge cases. They belong to the people closest to the customer, building within the product’s guardrails.

That belief is not new for me. It is the same one I had back when I was helping people click buttons: The person closest to a problem should be able to solve it, without waiting or begging on someone else to fix it for them. Ours tools have fundamentally changed. My rason for being has not.

That is what we are building at Dubot: a product’s own actions, turned into a governed layer any agent can call, ours, theirs, or one of their customers brought along. Dangerous writes should require confirmation before they happen. Every action should be logged. A real human being should be in the loop when the company decides it matters, and the agent can pick up the slack. Not because governance demos well, but because I have seen how far down the hard part goes. An agent loose in your product without guardrails it is a quick way to lose your relationship with your users.

Talk is cheap. We started Dubot to build the part that is not.