Antti Tevanlinna, Bosser
Interactive work is how you accumulate the control.
I often see people wanting to build autonomous agents. Somehow the early imprint of what is an agent has been this autonomous thing that just handles stuff. We picture agents in the cloud or some similar starry picture. I know I had this same notion in my head when I got started with AI.
I've been trying to get my agents working. Got some working really well too. I have one recurring and autonomous agent. The rest are interactive.
Interactive agents are such that you talk to them. They talk back. Kind of like a chat but not a chat. Those who know Claude Code, Cowork or Codex know this kind. We talk back and forth. Sometimes the agent sprints off and crunches for minutes and up to hours. But then we talk again.
I get the appeal. Autonomous sounds efficient. Full on automation. Infinite gains. But but: Infinite gains don't appear from thin air.
Interactive first. Always.
The right starting point is talkative. You pair with your AI. You use the AI as an assistant. Claude Code style. You watch what the agent produces. You correct it. You notice where it is brilliant and where it is a confident idiot.
This interaction is for both you and the agent.
For you: After a few weeks of this, something shifts. You start to sense, before you run the agent, whether a prompt is going to work. You start to know which tasks the AI eats for breakfast and which ones it chokes on. You develop a mental model of the capability and the failure modes.
That mental model is your asset - just as much the agent. The model you now carry in your head of what this technology can and cannot do on your specific work. You can try to deduce this knowledge a priori with careful guesswork and reading. Or you can just do the work and keenly observe how stuff really works. I can tell you: the former works. Latter just goes way beyond your working memory capabilities.
People want to speed up this learning. You cannot much. Your brain has to observe, think and absorb. You have to put in the hours.
For your agent: Nobody's model will work for your domain fully out of the box. You have to do the hours in teaching your agent your world. Sure: you can tell your agent: go learn our company. But the effective current way is that you teach your agent by curating the right context.
It turns out: you can create your agent rules and guardrails from your steering.
The ladder
I like to describe this as a progression in three stages.
Rung one: Interactive. You drive, the AI assists. Every important detail of the output goes through your hands. You're learning.
Rung two: Supervised agent. The AI drives a bounded task end-to-end, but you review the plan, artefacts and compliance before it ships. You're calibrating and you are more and more "away from keyboard".
Rung three: Autonomous agent. The AI runs the loop. You monitor and steer outcomes, not details of outputs. You are intent engineering. You are system engineering.
I know most of us want to get to rung three. The rungs before accumulate the steps to get there. Both for you and your agent. Both actors are kind of... learning.
Crawl, walk, run & the ladder
Try skipping the levels and make an agent that does proper heavy lifting without you in the interaction.
You don't specify the work precisely enough for the parts that matter. If you've never done a task interactively with the AI, you don't know which parts it handles well and which parts need scaffolding. Your agent brief becomes a wishlist with blind spots. The agent will comply, fill in plausible guesses and the whole of the output is like a minefield. Like a large intermediate level minefield from the old Windows operating system. Plenty of squares. Plenty of mines. All for you to gradually map out.
Interactive work looks slower. One person, one AI, one task. Feels like you're not scaling.
Interactive work is how you accumulate the control. How much human supervision does this task actually need. Once you have the control and real steering capability in interactive runs, you can decide what to automate. Before you know where you are in control, any automation is a guess.
Second, you can't evaluate everything as the volume is just too much. "Is this output any good" is a taste judgement. Sure, each sentence in the output can be rewritten with taste by you. Not sure if the LLM then added any real scaling.
Guesses may be expensive. The bigger problem is that guesses don't create learning and control unless you set yourself up for it. Supervised experiments are cheap. Set your experiments and runs to create you your guardrails, rules and your control system.
Third, all agents go wrong at times. When the agent goes off the rails, you don't catch as you don't know what normal looks like yet. There's no one in the building who's seen the AI do this task enough to learn. So you can't monitor.
The people I see actually scaling are the ones starting with the engineering mindset on rung one, then gradually moving to rung two with a clear-eyed view of what works. By month six they have real agents running on real work slices without the engineer interacting step-by-step. I still haven't seen anybody successfully start with the non-interactive for anything but simple email triage.
Why genuinely novel work may remain interactive
The autonomous agents are a class of agents and we will see much more of those. Right now, I see many trying to teleport there.
I run 80% of my work in really new and novel stuff. There is no way I can produce a spec for the thing I'm just trying to learn myself. The agent is a working companion. It produces some viewpoints. I critique. The agent critiques me. Bit by bit, something like the "right thing" emerges.
I believe this uncertainty, learning, and concepts being half-formed characterises real innovation. And I like to keep people in the real innovation. I'm still not hurrying up to do the autonomous agents in scale. I feel that currently is for the non-core work of mine.
But I keep a disclaimer about what I just said. I think you know why.