Is Meta-Design Here? and Is Attention Its Bottleneck?

Albert Shum asked a question in his LinkedIn article on meta-design  that I keep chewing on. For every hour of design, how many hours of work does your design agent do for you? He draws on Anthropic’s employment report, which apparently describes about six hours of agent work for every hour an employee works. It also says that 22% of interactive Claude sessions now focus on managing and guiding agent work, up from around 10% less than a year ago.

Shum calls this meta-design. The work shifts from creating the output to designing how the output gets created.

When I read that, I think I actually laughed a little. It sounded a lot like my last project.

More Output Than One Person Can Read

So, I was building an AI agent and writing its skills files, and that’s where my partial attention came back to bite me. The agent kept producing long responses and files, and almost every one ended with a question about what to do next. I answered quickly and only partially, with whatever I could process in those few seconds.

Looking back, I think I know why I was rushing. My days look really different now. I usually have several projects going at once, and each project has its own agent. On top of that, I have a few assistant agents that help with my own coordination and daily productivity. It sort of feels like managing a junior team of designers and researchers, except this team never sleeps. With a human team, I’d give instructions and then wait a few hours for an update. With agents, I never really have to wait. They come back within minutes, carrying work that looks like it should have taken a whole lot more hours.

And then they wait for me. It’s a bit like having a very quick worker who’s always standing at my office door, asking how they can help and what to do next. If I don’t hand over the next step right away, it feels like they’re queueing up for work. Somehow that makes me feel like I’m wasting time, as if I’m not properly resourcing their scope of work.

Before AI, I’d work on one project for a good stretch of time before moving to the second one. Three projects in a day was about the most I could manage. Now I probably move between three projects in two hours, and I switch back and forth, maybe six to nine times in that window. Every switch lands me in front of another agent that has finished something and wants a decision. I guess I really didn’t have the headspace to read any of it properly.

Shum warns that every design issue starts to feel urgent when an agent raises it. I think that’s true. I felt the pull to respond to everything, right away, and I’m not sure I made good calls.

Here’s the part that still bugs me. A while later in the skills files project, I gave the agent a very specific instruction. It replied that my instruction would conflict with three other things stated in earlier responses and files. That’s when it hit me. I hadn’t read or critiqued those files properly when the agent first created them.

And honestly, I can’t tell you how my rushed replies shaped everything that was built after them. I just don’t know.

“A shift from creating the work to designing how the work gets created.” – Albert Shum

To me, the scariest bit of meta-work is that my partial attention quietly became the instruction.

Creating AI Agents is a Design Task

AI Agent’s Skills files look like plain text, right? Just instructions written down. But while I was writing them, it felt more like designing a system. Each file sat somewhere in a hierarchy. Information had to travel from one file to the next. Knowledge sources had to connect to the right skills, and every little edit risked contradicting something that was already there.

I only got through it because I drew a schema. It showed me which files and instructions were prioritised, which were not, and where two of them would clash. Without the drawing, I’m pretty sure I would have been lost. I had to iterate on it several times, and the drawing was what let me see the whole thing.

I wonder if that instinct comes from somewhere deeper. I wrote before about the habits we build when we use our hands, and I guess drawing a system is one of those habits. It forces me to decide what connects to what. So maybe the craft that meta-design leans on is the craft of making a system visible enough to judge.

Shum asks where the next generation of designers will build the judgment to direct all this agent work. My project nudges me toward one possible answer. Judgment might grow from the habit of making structure visible. I only have one project as evidence, so I’m holding that loosely.

There’s also a bigger backdrop here. Many organisations are still in the construction/integration phase of AI. They’re re-laying the infrastructure and wires. And I’d argue someone with a design mindset should be sketching how that infrastructure and those wires connect.

Give Agents a Shape, and Give Yourself a Map

I’ve picked up three habits that seem to help. I’ll probably keep using them, though I’m still figuring them out.

Standardised outputs
The first one is to ask agents to report back in the same format. Think about how meetings work. There are minutes, decisions made, and action items. I found myself wishing every agent would answer that way. If each report came with a short summary, the decisions it made, the assumptions it leaned on, and the questions it needed from me, I could scan each one quickly and decode a lot less. I think a lot of my overwhelm came from context-switching between different shapes of output. A standard format would likely take that cost away.

Information Flow Schematic Diagram
The second one is to sketch the information flow before the agents start. For researchers and educators, this might be the most useful thing I can pass on. Draw how information enters, moves, and changes between agents and files. Then ask what each step assumes. It’s a bit like reading a map before a road trip. Surfacing assumptions on paper is much cheaper than finding conflicts three conversations later. AI also tends to sound confident even when something small has quietly derailed it, so the map becomes a decent check on that.

Slow Down
The third one is to reply in full or park the question. I used to feel bad when I let an agent wait, like I was keeping someone standing at my door. Now I’m trying to see a waiting agent as a normal part of the work. A half-answer sent in a rush probably costs more than a pause. This connects to something I explored in the present tense of AI. The agents set a fast pace, and I’m slowly learning that I’m allowed to set my own.

What Meta-Design Might Ask of Us

Shum wonders whether design teams will shrink, whether leaders will manage ten or more direct reports who each run their own agents, and whether we’ll spend more time reading and reviewing than making. I don’t have answers to those. I can only say what the daily experience felt like to me. The agents scaled easily. My attention didn’t.

So maybe the designers who do well in meta-design will be the ones who treat their own attention as the scarcest material in the system. They might design the reports, the schemas, and the pauses with the same care they give a layout.

What’s one thing you would ask your agents to return in the same format every time?