The new ChatGPT desktop UI makes it unnecessarily difficult to find regular chats. Chat, Code, and Agents feel like separate products occupying the same app, but the navigation doesn’t explain that model.
I’d make the modes explicit and keep Projects and Recent conversations visible:
💬 Chat | 💻 Code | 🤖 Agents Search
Projects
• Atlas
• Identity Ledger
• Home Power
Recent
• Finding ChatGPT Chats
• Harvester Research
• ...
One app, three clear modes—without hiding conversation history or making people troubleshoot where their chats went.
I keep thinking about what happens to design systems when the user is not only a person clicking through an interface, but also an agent trying to understand what it is allowed to do.
Traditional design systems gave teams reusable visual and interaction pieces. Buttons, cards, spacing, typography, states, components. But agents need a different layer of consistency: commands, source boundaries, privacy rules, evidence requirements, tool permissions, output contracts, and stop conditions.
That feels like the next version of a design system to me. Not replacing the visual system, but sitting beside it. A product should be legible to people and operable by agents.
Claude’s new reflection dashboard feels like a quiet UX signal.
AI tools are moving past “look what I can do” and into “look at how you’re using me.”
That’s the design problem I care about most right now: not just making tools more capable, but making sure they preserve human judgment while they help.
It connects directly to what I’m exploring with Atlas: behavior-first design, collaboration, and tools that shape attention instead of just producing output.
It’s been a while since I've been active in this space, but I want to start to think out loud again.
A lot has changed. I’ve been building, learning, making, rethinking what I want my work to be, and finding my way into a clearer version of the things I care about.
arXiv.org 
