From Copilots to Agentic Frontends: Notes on Lighting Talks at WorkOS
Tuesdays event focused on something I’ve been thinking about a lot lately: what happens when we move from chat-based AI to agentic systems that operate inside real workflows. Actual software that plans, structures, and executes work.
Here are the key ideas that stood out.
Agentic UX Is Not Chat
One slide captured it clearly: “When you move from chat to workflows, you need to give guidance.”
Chat is open-ended, eventually it looses the context. Workflows solve that contextual and structural problem.
The example shown broke a typical development flow into:
Questions → Research → Design → Structure → Plan → Worktree → Implement → PR
This is fundamentally different from a single long chat session. It’s staged cognition.
As designers, this changes our role. We’re no longer designing conversational responses or input fields. We’re designing guided systems for execution of specific tasks.
Controlled vs Open-Ended Generative UI by CopilotKit
One of the strongest frameworks presented was The Generative UI Spectrum:
Controlled – predefined components (charts, widgets). Agent selects what to show.
Declarative – combining pre-existing blocks (cards, lists, forms).
Open-ended – embedding full applications with maximum flexibility.
Controlled generative UI had clear advantages:
Easier developer experience
Pixel-level control
Secure by default
The tradeoff: higher coupling between frontend and backend, and the agent invokes components rather than generating them freely.
As someone who cares deeply about trust in AI systems, this resonated. Maximum flexibility is seductive, but controlled surfaces are safer and more production-ready. However there is still a question to answer - how would it be utilized by non tech users. How might we as designers and developers on the forfront of technology make that accessible to humanity.
Memory Is Still Fragmented
There was a discussion around OML (Observational Memory Layer) and the idea of a “master memory AI.”
Current systems:
Long sessions feel powerful
But context windows are unstable
Information fragments
We don’t build a rich, persistent picture
One provocative question stuck with me: “What if forgetting is a feature, not a bug?”
We treat memory loss as a failure. But intentional forgetting may reduce noise, drift, and hallucination over time.
Another interesting idea: instead of one agent with one memory, imagine an agentic set — multiple agents acting as one system:
Observer → Context → Buffer → Memory Creation
This feels closer to cognitive architecture than chatbot UX.
From OpenClaw to “TrustClaw”
There was a reference to moving from something like OpenClaw toward something more trust-oriented — think “TrustClaw” with hundreds of configurations.
The direction was clear: easier user friendly setup, structured flows for API calls, clear control over what the agent can access
This aligns with something I strongly believe:
If we want agentic systems adopted by non-engineers, permissioning and regulation must be first-class UX primitives.
The Two Phases of Agentic Frontends
So from robust agentic backend traind in completing specific workflows to the non tech user - front end and UX becomes the connective layer - human to life improving tech. One slide framed the future in two phases.
Now — Enablement: Building the tech that allows agentic frontends to exist.
One year from now — What’s next? Agentic frontends are ubiquitous. Then what?
This is the more interesting question.
Once every app can generate UI dynamically, orchestrate workflows, and execute tasks autonomously — differentiation won’t come from capability.
It will come from:
Trust design
Permission frameworks
Memory governance
UX clarity under complexity
My Takeaways as a UX Designer
This event reinforced something I’ve been exploring in my own work:
We are shifting from designing interfaces to designing bounded intelligence.
The hard problems ahead are not “How do we generate UI?” or “How do we chain tools?”
The hard problems are:
How do users regulate what the agent knows?
How do we visualize agent state and memory?
How do we prevent silent drift?
How do we design connection?
How do we make autonomy functional yet regulated?
Agentic UX is not about adding AI to screens. It’s about designing systems where humans remain in control — even when the system can act independently. And that’s where real product strategy begins.
