Surfaces
Pattern 21 of 26
Chat Interfaces
Looks simple. Nothing about it is simple.
ChatGPT, Claude.ai, Gemini, and DeepSeek are where most people first encounter what agents can do. The format looks simple because the interface is simple. The engineering underneath is not: streaming responses, progressive rendering, multi-turn context management, rich output formats. Chat is the lowest-friction entry point and also the most constrained one.
Why it matters
Chat sets the mental model most people carry into every other agent surface. Whatever they believe agents can and cannot do is shaped here first. That is a lot of weight for a text box to carry.
Deep Dive
The chat interface looks simple because it is meant to. You type, the model responds. But the implementation is not simple at all. Streaming responses require the frontend to handle partial text, render progressively, and update without layout shifts. Multi-turn context management means carrying the right information forward without silently dropping earlier turns that the model still needs. Rich output like code blocks, tables, and math notation requires a rendering layer that goes well beyond displaying plain text. None of these are hard problems individually. Doing all of them well at the same time, reliably, is.
Claude.ai shipping artifacts in June 2024 with Claude 3.5 Sonnet was a useful design decision to watch. The insight was that some model outputs are not conversational. They are documents, apps, or code that the user wants to inspect, edit, copy, and run. Putting those in a side panel separate from the conversation was an architectural choice that reflected that distinction. ChatGPT's Canvas followed the same idea. The pattern is now widespread: separate the conversation from what the conversation produces.
The structural limitation of chat is that it is sequential by default and it puts the burden of describing context on the user. I notice this every time I use a chat interface for something that requires more than a few back-and-forth turns. The description of what I need is always slightly wrong. I miss details I did not think to include. The model works with what I gave it. Generative UI and multi-agent workspaces are the directions this evolves toward. But chat is still the right starting point for most users because the friction to begin is low and the expectations are calibrated to match what it can actually do.