I dictated a presentation brief to Sky whilst on the train. By the time I got to my desk I didn't want to be in a chat anymore. I wanted to be in PowerPoint, dragging slides around, wordsmithing sentence by sentence. Something different happens in the problem solving process when you're moving the boxes yourself. 40 minutes of work and a few meetings later, I handed the work back to Sky — same agent, same task — with a note that I'd made changes and wanted a review pass. The review came back tight. I let Sky finish the cleanup without taking the file back.

One task. Three interfaces. A gap in the middle. Me driving for part of it, Sky driving the rest. The only thing that stayed constant through all of that was the task. Everything else moved around it.

The task is what makes the handoff possible. Without it, every channel switch is a context reload, and every gap is something I have to stitch back together. The task gives the work a place to live that isn't tied to who's currently holding it or what app they're in.

I’ve always wondered why when using AI products for work, it was good but it never seemed to… flow.

My conclusion is that the human and AI interaction just wasn’t fit for how I work. The default AI interface is a chat. Every interaction is its own conversation, and every conversation is its own silo. You can try Projects, but the moment you sit down to make one you're deciding what counts as a project and what doesn't and where the boundary is. Or you hope the AI's memory function pulls the right thread whenever you ask a question. It might. It might not.

Ultimately, conversations are how I think. They aren't how I get things done.

So I started asking a different question. What's a better shape of work for AI to slot into?

Looking at how anything actually gets coordinated — at home, at clients, with my team, the unit they all share is the task on the todo list. The smallest stable thing that multiple people can hand back and forth without losing where they are.

I took this and built it into Sky System to test it out. I drop tasks into it, agents pick them up, and the work happens against the task rather than against a chat thread. That on-the-train PowerPoint work is one of many tasks I’m multi-tasking with AI. I do the parts humans do best, AI does the parts AI does best. All without losing context and coordination.

The screenshot above shows a snippet of what human and AI todos in Sky System looks like at the moment. You’ll see that it’s a standard todo list interface. But what you might notice is who is owning the task. Some of them is assigned to me, some of them is assigned to an agent waiting for me to provide a review. Below is another screenshot of how to assign the task to the agent best suited to do the task.

But why tasks? Because real work isn't a single exchange. It involves multiple inputs, waiting periods, shifting context, back and forth across time. A chat isn't designed to hold all of that — it's built for the moment, not for the gaps between moments.

And tasks are. A todo sits in a list and waits. It holds its place for whoever picks it up next, whether that's you tomorrow morning, an agent in five minutes, or a colleague after their next meeting. The work stays put even when no one is actively on it.

Below is a screenshot of the task created for previous article’s LinkedIn post. A bunch of instructions, with context from various points, and a few back and forth exchanges with me. I then did the rest on LinkedIn itself.

Something else changes when work lives in a task. The AI agent's relationship with tasks shifts. In a chat, it’s asked a question, the agent answers. Against a task, the agent has assigned responsibility to come back to. It can see what's been done and what hasn't. It can report progress, ask for input, mark something resolved. The unit of work gives the AI agent a goal and objective.

We've known this for humans working with humans for a long time. It's why we have tickets, sprints, lists on the back of envelopes. We just hadn't built it for humans working with AI.

How humans and AI meet to do work matters as much as what either one can do alone. The interface decides whether collaboration actually happens. When it doesn't fit the work, we see powerful technology stuck behind friction. But when it does, it feels like a gust of wind at my back.

This isn’t a one-size-fits-all problem. The system I built works for me because most of my work relies on multiple people with a lot (probably way too much) context switching and parallelism. I need a way to put work down without losing it, and pick it back up hours or days later without having to reconstruct where I was. In this setup, the list and AI carry the work, and I get to do the thinking and shaping.

Someone whose work is mostly focused work, or who finishes things in one sitting, would design something different. They might be perfectly served by chat. There's a class of work that probably should stay chat-shaped: pure thinking out loud and creative exploration before you know what you're making.

It's this sophistication that necessitates the interaction between human and AI to be deliberately designed to fit the job that is being done.

This is Sky System article #2. The first one, "I'm watching my agent watch me", is here. More to follow.

Sky System (Sky) is the bespoke AI multi-agent stack I'm building for myself. Part productivity tool, part live experiment in how humans and AI actually collaborate. This series is the running notebook: what's working, what isn't, what's surprising me.