Because Text Is the Universal Interface, and as more applications are driven by AI agents (an LLM called in a loop to take actions to complete a goal), we’re going to see more APIs turn into chat interfaces, where you’ll ask the API a question or give instructions in natural language.
Oh no! My service is already too slow, now I have to wait an extra 2-20 seconds for the LLM to write a reply?
It’ll be OK, because the future will be LLM agents all the way down. You’ll give your AI personal assistant a task, and it’ll talk to a horde of other specialized AIs to get the job done.
This is a good thing, because we can stop “pretending”:
AI Written, AI Read cartoon - Marketoonist | Tom Fishburne
Here’s how this might progress:
Proof of Concept: Forms
A simple example to motivate how you can use chat to actually improve (!) the process of collecting/producing structured data: Instead of using tedious form builders (e.g. Google Forms), you can simply describe what you want (“Startup waitlist form with name, email, company, and job title”) and have the schema inferred for you.
This makes it trivial to include logic that would be annoying or impossible in traditional forms (“If they’re any kind of software engineer, ask for their GitHub”), with custom validation (refuse to admit obviously fake emails), and the form responses remain structured according to the inferred schema for easy analysis.
There’s already Formless by Typeform, YC’s Raz and Surface Labs, and (my) open-source TalkForm AI.
In progress: Customer support agents
DoNotPay uses a chatbot to negotiate your bills. Meanwhile, companies increasingly use chatbots (through e.g. Intercom) for customer service, leading us to chatbots negotiating with chatbots.
Soon: Agent specialization
Today, if you’re building an agent that needs up-to-date info (past the model training cutoff date), you can give the agent browser access (e.g. with SerpAPI) to let it search for itself, or ask a specialized service like Synapse. For example, you might have a coding agent that needs up-to-date documentation.
But this breaks down into many hard sub-problems. Quickly and effectively searching/summarizing code documentation is quite different from getting a medical diagnosis, which is quite different from picking a restaurant, even if they all are RAG. So, we will likely see more specialization: your Coding Agent will ask its questions to someone else’s Documentation Agent, and your Party Planner Agent will ask questions to the Restaurant Reservation Agent which will talk to your Personal Preferences Agent.
Later: API for your personal preferences
The combination of constant data collection (Rewind AI, Tab, Humane) and AI clones (Character AI, Delphi) allows you to create a Horcrux, your digital twin, the finally convergence on a long-awaited universal personal representation. The real power here will be composing it with the other Agents. The clones will manage privacy and expose tiers of access to other applications, allowing every application to become perfectly personalized.