What is an LLM Agent for non-technical people?
In our ongoing to series to discover what the future of automation might look like, today we’re taking a deep dive on Agents for all you non-techies out there. Even though this subject area can get technical very fast, we wanted to present a simple, easy to understand introduction to LLM agents.
First, some basics. An LLM (Large Language Model) is a form of AI that uses natural language processing techniques to generate meaningful text that often comes in the form of dialog. This is the engine that powers ChatGPT.
Agents are an extremely new development at the forefront of this technology which represents the next stage of the LLM capabilities. As users, we’ve become accustomed to prompting the LLM with some text to generate an appropriate response. Agents take things a step further and allow the LLM to take actions on behalf of the user (e.g. schedule a meeting for you, like Siri or Alexa).
It’s easier to start explaining what an Agent is by first listing a few of the things it is not.
An Agent is not:
- A chatbot
- A simple LLM (by itself)
- A person (yes really)
Agents are a combination of advanced programming and prompting techniques which allow an LLM to reason about the world around it including any context you give it, and take actions to accomplish goals you set for it. It’s like an advanced version of Siri or Alexa that can not only set meetings when you tell it to, but also engage in fluid conversation and proactively manage your calendar, alter spreadsheets, interact with software, and generate content in response to the conversation you’re having with it.
It’s best to understand an Agent’s capabilities with a few examples.
Take the following instruction:
“Book a meeting tomorrow at 09:00 if the client is available, if she cancels send her a follow-up email and update the records in our CRM.”
The following will occur:
- The agent will understand from your conversation history that the “client” you are referring to is the (fictional) email `jane@my-client.net`
- The agent will then check Jane’s calendar and see if there is a slot open for 09:00 and try to book a meeting.
- If Jane cancels, the agent will draft a full follow-up email trying to book another date
- Finally, the agent will update the details of the expected meeting in your CRM.
The tasks you give the agent can also be even more abstract:
“Develop a content strategy based on the best-performing posts by engagement across our LinkedIn, X, and Instagram channels for the next 2 months, it must include influencer marketing”
In this case, the agent would go ahead and read all the metrics across your media channels and identify those with the highest engagement. It would then use those to form a base from which it builds a content strategy that it lays out in Excel (or Google Sheets, we don’t judge). Finally, based on the posts and your brand voice, it would search the web for the best influencers to partner with as part of the strategy.
You didn’t tell it to search the web, it reasoned that it had to do so (in the absence of specific instructions) in order to identify the influencers that would be right for you. You get the picture.
With the examples out of the way, we can define all the ingredients for our lovely agent soup. An Agent is an LLM augmented with the following capabilities:
- Long term memory
- Tools to use (e.g a calendar or a web browser)
- Logical reasoning
- Planning
Why is an Agent different from a regular LLM?
A regular LLM doesn’t have the capabilities listed above, namely long-term memory, planning, tools, and reasoning. An LLM has a context window (a maximum word count for its inputs and outputs) and that’s the extent of its memory! Agents on the other hand significantly extend what’s possible by augmenting the LLM with supporting technology.
LLMs also cannot use tools like regular software or web browsers, because it has no way to interact with the environment or produce an output other than text. Agents can engage in conversation, extract what’s relevant, and use that as input into other software.
Moreover, an Agent can be “taught” to use any tool! Whether it’s a web browser, Outlook, Google Calendar, or your favourite messaging application like Slack, Teams, Whatsapp, or Telegram.
Reasoning is what ties this all together. The agent will attempt to reason about the task being asked of it by chaining together “thoughts”. Without getting technical, “thoughts” enable the LLM to inspect its own output and decide if it is relevant or useful to the task at hand.
Planning is a fancy word for letting the LLM “think” ahead and decide on a series of steps it will take to accomplish the goal, instead of focusing only on the immediate next step.
Agents are the next enterprise platform
Given the amazing capabilities Agents have demonstrated we believe they will fundamentally change the way we work in the future. As shown in the content strategy example above, Agents are a first-of-a-kind technology with the ability to autonomously complete complex tasks from end to end. Although the technology is still nascent, it has the potential to change the way we think about work and coordinate large-scale knowledge-based work in the future.
That said, there are still some limitations we came across when building these in-house. We found that for the best performance, an Agent should specialise in a given domain (e.g. Marketing, Sales); even better if it can niche down on a vertical like Finance Marketing or Tech Sales.
Just like people, specialisation can help ensure that the Agent produces results with a consistent quality that’s perfect for your business. More generalised Agents can get the job done but can be unspecific in their approach, this can affect the quality of any generated content. Just like ChatGPT, non-specialised Agents can get by in terms of general knowledge, but their performance drops off when you start to ask it for specific details which might be niche to your business.
How can I use LLM Agents in my business?
If you’re asking yourself this, you’ve come to right place!
At Solitude, we’re building a platform to connect Agent builders with innovate enterprises who are looking to automate more of their business than ever before. For enterprises, you can think of the Solitude as Upwork or Fiverr for AI employees that can carry out any task, which you’re free to hire on a monthly basis.
Builders, can use our platform tools to develop specialised agents for any job role including Marketing, Sales, Customer Service and Finance to name a few. Agents can build for any functional area, including ones we have listed here like Design!
If you found this an interesting read and want to learn more, we’d love to extend an invite to you via our early adopter’s program which you can sign up for at our website: https://solitude.com
Sources
- https://lilianweng.github.io/posts/2023-06-23-agent/ — for more technical audiences
- https://www.forbes.com/sites/alexanderpuutio/2024/03/22/what-ceos-need-to-know-about-the-next-frontier-of-ai-ai-agents/ — a Forbes article on agents
