What is an AI Agent
What is an AI Agent? Is it VC fluff?
An AI agent is an autonomous system designed complete a specific task or achieve a specified goal without human intervention.
1. How is this different from a chat based LLM?
A human wants a specific outcome. He promtps the AI in a chat interface to get that specific outcome. This could be a multi step process. For example analyzing data, then extracting it, looking for a specific signal and then creating a report based on that.
In the case of an agent, IT has been given the outcome. It operates independently, making decisions and taking actions based on its programming and environmental inputs. Process information and determine the best course of action to achieve their objectives. They work towards specific objectives, which are predefined. The agent can improve its performance over time through experience and feedback (aka 'evals'). These are multi-step processes as opposed to a single step which the chat based LLM does.
The agent has access to real time specific data, where as the Chat LLM just has the data that it was either trained on or prompted with.
2. How is it different from a vertical Saas app?
I'm going to assume a Saas app which is used for out-bound marketing emails. Using a traditional approach this is what the flow would look like.
If this process was done by with the help of agent, this is how it would look:
The agent is programmed to control the traditional Saas application, make decisions based on its reward function, and take actions to achieve specific goals without human intervention.
A human could use this as a tool, refine the agents process at every step by correcting its mistakes, suggesting what could be done better (such as signals for finding better leads, or better language to be used in drafting the emails).
But in the case of an Agent, which is trained with a predefined objective, the agent would have context on what the final outcome is, and what it should do in order to get to that outcome.
For example, if someone had to regularly do the task of going to a website, downloading a report, then analysing it, looking for some specific fields of data and then processing that to share it with their team.
This process could be done with an AI Agent, which would have access to the browser, it would go to the website periodically, download the report, then perform the analysis based on the output that was needed, create the report and share it with the team.
3. Is this just hype?
Like most things in the AI space which are backed by VC firms, there is a lot of hype involved. Tools like Devin have had massive failures on launch. But at the same time, there is super fast growth and cases where there are huge business opportunities to be built using these agents.