From chat window to bot swarm: five stages, each learned when it broke

X (@g0mu_)

My path with AI went through five stages, and I only understood each one when it broke.

1. A chat window. I asked, it answered, and I copied the answer somewhere. I was the agent. The AI was a very smart text box, and every step between the question and the result was me.

1. Agents. Software that takes steps on its own: it opens files, runs code, checks a page, and reports back. For the first time, work finished while I was doing something else.

1. Agents talking to agents. One agent builds and another checks it. On my floor today, a QA agent opens the live site and sends back screenshots of what actually shipped. My problems moved from prompts to handoffs: what does the next agent actually know?

1. Loops. Agents that run on a schedule without me asking. An overnight builder ships while I sleep, and a brief lands in the morning. This is where I learned that a report saying "done" and a thing being done are two different events.

1. A swarm. Many agents, many schedules, one person. I'm still in this stage, and the lesson so far is boring: the more agents you run, the more your job becomes deciding what counts as proof.

Each stage left me less building and more checking.

#MakingSandThink

LinkedIn

My path with AI went through five stages, and I only understood each one when it broke.

It started in a chat window. I asked, it answered, and I copied the answer into whatever I was working on. Looking back, I was the agent. The AI was a very smart text box, and every step between my question and a real result was mine.

Then came agents: software that takes steps on its own. It opens files, runs code, checks a page, and reports back. For the first time, work finished while I was doing something else.

Then agents started talking to each other. One agent builds and another checks it. On my floor today, a QA agent opens the live site and sends back screenshots of what actually shipped. At that point my problems moved from prompts to handoffs. What does the next agent actually know, and what got lost on the way?

Then loops: agents that run on a schedule without me asking. An overnight builder ships while I sleep, and a brief lands in the morning. This is where I learned that a report saying "done" and a thing being done are two different events.

Now it's closer to a swarm: many agents, many schedules, one person. I'm still in this stage. The lesson so far is boring. The more agents you run, the more your job becomes deciding what counts as proof.

Each stage left me less building and more checking.