I was writing a piece of communication work for Tidra, and I was doing it with Claude. Not the way people imagine, where you type one line and something finished appears. The real way. I would draft, it would respond, I would push back, it would try again. Back and forth. The way you would think out loud with someone whose judgement you trust, across a table, working towards something neither of you could see clearly at the start.
For a while that was exactly right. I was still figuring out what good looked like. Each turn taught me something, and the conversation was where the thinking happened.
Then, somewhere in there, I noticed the conversation had stopped being a conversation.
Two different things, named plainly
Before I go further, it is worth being exact about what I was actually doing, because the two options here are often blurred together.
The first is what people mean by working with a skill. You sit with a general AI model, Claude in this case, and you work conversationally. You bring well-crafted instructions, you read what comes back, and you refine it turn by turn. You are present for every round. The intelligence is in the model, but the direction is in you, live, in real time. Nothing runs without you.
The second is a built agent. This is a custom workflow made for one job. The steps are wired in ahead of time. The tools it can reach for, the order it works in, the decision logic it follows when it hits a fork: all of that is decided in advance and built into the system. Once it exists, it can produce the output on its own, without you sitting there for each round.
One is a live conversation you stay inside. The other is a process that runs whether you are watching or not. That is the whole difference, and it matters more than the word “AI” sitting on both sides of it.
The moment it changed under me
Here is what actually happened at my desk.
Early on, I was discovering the answer through the conversation. By the fourth or fifth piece, I was not discovering anything. I already knew what good looked like. I could tell within a line or two whether a draft was going to work. I knew the moves I wanted, the tone I was after, the things to avoid.
And I noticed I was doing something slightly absurd. I was typing the same instructions again. And again. Reminding the model of things I had reminded it of an hour before. Wishing, quietly, that they would just stick.
That wish is the signal. When you find yourself repeating the same guidance across sessions, hoping it holds on its own, the conversation has already done its job. You are no longer thinking with the model. You are re-entering knowledge you already have, by hand, one turn at a time.
That is the moment a skill starts asking to become something more structured. Not because the agent is the superior tool. Not because conversational work is a beginner’s phase you graduate from. But because the task itself had quietly changed shape underneath me. It had stopped being an act of discovery and become an act of repetition. And repetition is precisely what a built workflow is for.
What the fork actually tells you
I think a lot of the confusion around this comes from treating it as a question of sophistication. As if building an agent is the grown-up version and conversing with a model is the training wheels. It is not that.
The conversation is the right tool when you are still learning what you want. It is where judgement forms. You should stay in it as long as the task is teaching you something, and you should not be in any rush to leave.
The built agent is the right tool when the judgement is already settled and the work is now about doing the same known thing reliably, without you present each time. The value of building it is not speed for its own sake. It is that your attention is finite, and there is no reason to spend it re-typing decisions you have already made.
So the useful question is not “should I build an agent.”
It is simpler than that. It is: “am I still thinking here, or am I just repeating myself?”
While you are still thinking, stay in the conversation. The day you catch yourself wishing the instructions would stick, the task has told you what it wants to become.
I did not decide to build an agent because it was the advanced move. I built one because the conversation had ended and I was the last to notice.
This was not only a Tidra story
It is a fork we sit inside with clients constantly, more often than either side expects going in.
Someone comes to us still figuring out what they want. A voice, a message, a piece of thinking that has not landed yet. We work with them the same way, conversation first, because that is where the judgement gets formed. Then, at some point, we hear it: the same correction, said a slightly different way, for the third or fourth time.
That is the same signal, showing up in someone else’s work instead of our own. Not a sign the conversation failed. A sign the task has already told them what it wants to become.
We do not walk in assuming which side of that line a piece of work is on. We have just felt that particular repetition often enough, in our own work, to recognise it the moment it shows up in someone else’s.
