The Five Eras of the Chat Interface

The chat window looks the same as it did in 2010: a box, a cursor, a send button. What sits behind it has been rebuilt from the ground up five separate times. Each era solved the previous era's ceiling and introduced a new one.

Era 1: The decision tree

Press 1 for billing. The earliest bots were phone menus in text form. Cheap to build, infuriating to use, and blind to anything the author had not anticipated. Their ceiling was imagination: they could only handle questions someone had scripted in advance.

Era 2: Intent classification

Machine learning let bots map many phrasings to a single intent. "Where's my stuff," "track order," and "did it ship" all routed to the same answer. Better, but still a lookup. The bot matched you to a bucket and read from it. Step outside the buckets and it fell over.

Era 3: Retrieval

Large language models arrived and bots could suddenly read your help docs and answer in fluent prose. This is where most "AI chat" on the web still lives today. It is genuinely useful and genuinely limited: it can tell you the return policy, but it cannot process your return.

Era 4: Tools and actions

The current frontier. Give the model access to your systems and it stops describing actions and starts taking them. It reads the order, issues the refund, books the slot. This is the leap from answering to doing, and it is what today's agent platform is built around.

Every era kept the same interface and changed the mind behind it. The box did not evolve. The thing typing back did.

Era 5: The ambient agent

Next, chat stops being a place you visit. The agent lives wherever the conversation already happens: your phone line, Slack, WhatsApp, a video call, another agent's request. The window dissolves into the channels. That is the bet behind letting one agent deploy everywhere at once, and it is closer than the static chat box makes it look.