No Dishwasher, Running Agents

I was on a call with a friend yesterday. We were talking about a thirty-minute autonomous build I was running in the background of our conversation. The agent was planning, building, evaluating, and looping until it hit its goal. We were watching the log roll past in another window while we talked.

I got up to refill water. I broke a glass. I stopped the call to clean it up. While I was on my hands and knees with paper towels, I noticed I was going to have to wash the dishes in the sink myself, because my apartment does not have a dishwasher.

I told my friend. He laughed for a long time.

The framing he landed on, after he finished laughing, was: the future maid is not going to be human, but the dishwasher is still going to be a sink and a sponge. That is correct. That is approximately exactly how it is going.

The frontier and the kitchen are in the same apartment

Most of the journalism about AI is written from one of two angles. Angle one is the lab angle, where the writer is interviewing a CEO who lives in a world of compute clusters and policy debates. Angle two is the impact angle, where the writer is talking to a worker whose job has changed or vanished. Both angles are correct. Neither one captures the texture of how the people I know actually use these tools, which is from a small kitchen, in between unloading the dishwasher we do not have, while a calendar reminder tells us we forgot lunch.

The frontier and the kitchen are in the same apartment. The agent is running in the same room as the laundry. The model is generating a thousand tokens a second next to the cat I do not have but my dog Francis is fine, thank you. The future is already here, and the future is doing dishes by hand.

What an AI agent workflow actually changes about the work

What it changes is the relationship between you and your time.

Before the agent, my time was spent on the things only I could do, plus the things I should not have been doing but did anyway because nobody else was going to. Drafting the email I had drafted forty times before. Pulling the list. Running the report. Cleaning the data. The agent took the second pile.

After the agent, my time is spent on the things only I can do, plus the dishes. The dishes did not get faster. The dishes are still hand-time. What changed is that the second pile, the should-not-be-doing-but-did pile, fell off the side of the desk. The agent does it. I check it. I move on.

That is the sentence that is hard to write without sounding triumphant about a tool. Let me try again. The thing the agent did for me was give me back the late afternoons. The late afternoons used to belong to the second pile. They now belong to me, or to my partner’s calendar, or to the dog, or to a dish.

What agentic AI actually is, from the kitchen

The thing that gets lost in coverage of autonomous AI agents is that they are not magic. They are a class of software that can plan, execute, and revise without a human clicking approve at every step. Agentic AI is what happens when the model gets a goal instead of a prompt. The model figures out the steps. The model runs them. The model checks its own work. This is what I mean when I call the system I run agent-native. The human is not out of the loop. The human is at the checkpoints, not the keystrokes. I got out of the loop for the thirty-minute build, not the outcome.

The content engine I run at Plumwheel is built the same way. One recorded conversation becomes weeks of marketing because an agentic workflow handles the steps between the source conversation and the published piece. That system is what I wrote about here. The dishes and the agent are both real. They coexist. The agent doesn’t do the dishes. The dishes are still mine.

The unromantic truth

The unromantic truth is that the dishes will keep being unromantic. The agent will not unload your gym bag. The agent will not call your mother. The agent will not pick the nice cheese for the party. The agent is excellent at the second pile. The first pile is still the first pile.

If you are looking for a sign that AI is not actually about to replace your whole life, the sign is the dishes in your sink. The model can tell you the order of operations. The model can give you a four-step sponge protocol. The model cannot pick up the sponge. The picking up of the sponge is, for now, the human’s job.

I find this comforting in a way I did not expect. The frontier is wild, and the dishes are still mine, and both things being true at the same time is what makes any of this feel real.

If you are a founder or early-stage team trying to figure out how to actually build with agents rather than just demo with them, RAGnos Labs is what I built to do that work.

Cheers.