Human-Agent Interaction Is Becoming the Bottleneck
I’ve been a heavy user of agents for a while now — not just coding agents, but agents in general, for all my work. And I’m beginning to bump up against some walls when it comes to human-agent interaction. I want to share some thoughts on what those walls are and what might actually solve them.
The Chat UI Has a Low Ceiling
Look at the direction agent UI is going. Cursor 3.0 just shipped, and its big selling point is managing tons of agents — local or cloud, running in parallel — all in one UI. Claude Code looks similar. And it’s not just the agents: a ton of SaaS products (Linear, PostHog, Attio) have converged on the same ChatGPT-like interface: chat front and center, a sidebar of previous conversations.
This kind of UI has a low floor — it makes it easy to get started with agents. But I’d claim it also has a fairly low ceiling. Once you do a lot of your work (or even all your work) through agents, you hit a wall.
The wall is sensemaking — making sense of all that work after it’s done.
Narrating Your Work
For the longest time, I’ve been a fan of this concept called narrating your work. The idea is simple: you keep synthesized notes of the things you’ve been working on. A work log.
I was first exposed to this at Google, where engineers posted weekly “snippets” — short updates on what they’d done that week. Most people dropped a few bullet points. But every once in a while, there were those three or four engineers (you know who you are) who wrote mini essays in their snippets. Subjective, synthesized views of what they were working on, what was working, what wasn’t, what their plans were.
I got enormous value from reading those. These were engineers in distant parts of the company — it was one-sided in the sense that I never worked with them or met them in person. But reading their essays gave me a deep insight into the kinds of work they were doing. It made me feel a kinship with them. And it’s what encouraged me to keep up my own snippet-writing habit.
The primary audience is yourself — reflection over the mountain of work — but the secondary audience is everyone else who stumbles across it.
The Sensemaking Problem
Here’s the point: all these agent UIs make it really difficult to make sense of all the work you’ve done with your agents. You cover ten miles in a day where you would have covered one without agents. But at the end of the day — or the week — you’re left disoriented. You’ve done so much work. But if you try to think back: What did I actually accomplish? What’s the higher-level picture? It’s hard to reconstruct.
You need that synthesis so you can learn from it and plan your next chunk of work. And that, I think, is the main shortcoming of current agent UIs. They let you do a ton of work, but leave you in a disoriented place. Synthesis over all that work is an unsolved problem.
Agent Kitchen: A First Attempt
So I built a tool called Agent Kitchen that gives me a higher-level view of all the work I’ve been doing with my agents — across repos, across my home directory, across random directories on my computer. I wrote about the motivation in a previous post.
I’ve been using it daily since I released it recently. It’s now my primary interface into all my Claude Code sessions. I get a global view of sessions across different repos and directories, and I can click to resume any of them.
But even with this, I want more. This still isn’t quite the high-level synthesized narrative I’m after. Ideally, I’d want a paragraph or two of plain natural language at the top — a synthesis of what I got done in a given repo or directory, with links off to the relevant sessions so I can click to resume them.
Another common annoyance: agent chats grow really long, with many turns between you and the agent, and you just have to scroll through them. There’s no minimap, no bookmarks to just your inputs — nothing like the minimap in VS Code that gives you a global view of a long file. These are small UI issues, but they add up to the same problem: agents treat everything as a wall of text.
UX Is Now the Bottleneck
Overall, I think we’re arriving at a place where the UI/UX of using agents is becoming the bottleneck in how effective we can be with them. The models are getting better. The harnesses and agents are getting better. But I believe we’re now bottlenecked by the UX paradigm.
We spent the last two years making agents that can do more work. Now we need interfaces that help humans understand all that work. Until then, we’re sprinting blind.