Learnings on AI agents

by Sheng Yi, 01/17/2024
We began developing Language model-based AI Agents in June 2023. Here are a few things we learned so far.
First, a little background on AI agents. It is software (which can be embodied in hardware) to perform tasks on behalf of humans. Many people imagine a future where AI Agents do most of the work, leaving humans free to explore new and possibly more exciting things. The recent advancements in large language models have brought this vision much closer to reality, sparking excitement among its advocates.

Search is not consulting and passive chatbot is not agent

Like many startups, we began by developing chatbots for consulting in various verticals, using retrieval augmented generation (RAG) with GPT. Despite the increased competition due to OpenAI's GPTs, we have identified deeper issues inherent in the product itself. Our customers are saying “Asking questions is hard” when consulting with our finance chatbot.
It is because consulting is not search and chatbot is not agent. The key difference is during consulting people do not need to worry about driving the conversation to achieve the consulting goal. This is an agent rather than a passive chatbot. To build such a consulting agent, you need to plan the workflow of consulting on top of indexed knowledge points. However such workflow planning data are rarely digitized before.

People prefer delegating than learning if possible

When talking to users, often we hear “Can you just do it for me? I do not want to learn and then do it myself”. Aside from self-improvement where learning is the primary aim, many learning activities are undertaken to solve problems. In those cases, people love shortcuts to their end goals.
AI agents can provide such shortcuts by autonomously navigating the space of actions. They eliminate the need for people to learn and practice what agents are already capable of. And this is game-changing.

AI agent for web work

Over the past few decades, with the advent of the web and cloud technology, many user problems can now be solved online, and many more are transitioning to the web. Now a few research and product teams including us are looking into AI Agent for delegating the web work.
Similar efforts for App
I believe that in the next decade, people will shift from using apps or websites and instead delegate their work to agents. Simplegen will release in Feb, an AI agent for web works like “set up mail out of office message” or “submit job applications on Linkedin”. Sign up for early access here and we are looking forward to your feedback.