Welcome to a new thread
CES 2025
AI agent well in the Nvidia CES 2025 conference Jenson has stated the Artificial Intelligence or AI roadmap for the next 5-20 years probably.
AI agents are not so-called new concepts in very easy words it's a robots that don't look like robots and perform specific tasks by talking help of peer robots/agents to accomplish goals.
AI agents are the task performers who understand, evaluate and perform tasks to achieve goals with step-by-step process with logical reasoning for each step.
Let's understand it better with real-world use-case
First prompt: Give me react code component for simple todo app with to-do input and submit button
chatgpt: perform the task and return the code output
Second prompt: In the same codebase, add a method to create tasks and list down the task below the form.
chatgpt: perform the task and update the code output after adding a list of tasks and a method to add new tasks
Third prompt: Continue by adding a delete button with each task on the right side of the name of the tasks
chatgpt: perform the task and return >the code outputWell, one can go on and on and after getting the satisfied answer you can copy the final code in vs-code and get the app running.
What if a GPT or LLM mode itself takes care of these tasks and adds the required methods and UI components step-by-step on its own and that will be called an AI agent? A robot that basically does things for you but doesn't look like a robot.
How to make this stuff, AI agent?
Before that understand the basics
Using first principle thinking one can understand that it's a combination of again input and output.
One task can be divided into multiple small tasks and each task will return the output that goes to the input to perform the next immediate task until the last task is achieved or the goal line is reached.
Think of it like a linked list node, one node has a value and address to the other node and the other node has a value and address to the other node and in this fashion, all the nodes are connected making a so-called linked structure.
In every step, the LLM model or GPT will evaluate, think and perform the task.
Example
For example, let's suppose I want to create an AI agent that will generate a code repository for me with all my basic everyday packages installed such as next.js, shadcn tailwind CSS and other libraries like Firebase and supabase, or CMS
Now how I can create an AI agent to perform these tasks
The first thing is to break it to as much as possible as you can to understand the nitty-gritty of the situation.
- Create a Next.js repository
- Install Tailwind CSS and Shadcn library
- Create my repository structure including naming convention and coding preferences
Our LLM model will perform each task in a step-by-step fashion and if an error occurs in b/w it will not skip unless we instruct the AI agent to skip (which we generally do)
AI agent performing STEP 1
In the first step, the LLM model or AI agent will only look for a way to create a next.js repository by running the yarn add next react command, pages directory, next.config.js file
AI agent performing STEP 2
Once that is done and the app is running, we can instruct the LLM model to move to the next step and install 2 packages tailwind.css and shadcn UI library. Again the LLM model will perform a task in synchronisation by installing Tailwind CSS, adding the tailwind config file and then it will move to create the components directory to add the shadcn UI library
AI agent performing STEP 3
Once that is done, the same output will go to the LLM model to perform the last task that creates the repository with a naming convention and coding preference which the LLM model can either ask from user input directly, such as creating a pages directory, add src directory and modules directory with Navbar, Header and Footer components using shadcn and tailwind CSS.
How to build one?
This concept is more related to chain-of-thoughts prompting.
Providing chain-of-thoughts along with the process of the LLM model helps him to be more logical in reasoning and perform better tasks, reduce ambiguity as well as create more expected results.
Langchain frameworks help to create this chain of thought prompting along with state /context management with some LLM models' own memory space that works alongside the entire agent.
We have other frameworks as well as OpenAI API, Claude API, Gemini, Hugging Face for LLM models and creating AI agents.
Why AI agents are trending and the next BIG thing?
AI agents are the starting point and the goal is to reach AGI
Meaning computers can be run using robots or computer robots.
All the repetitive tasks on the browser can be performed using AI agent or automated
Designing the same designs on Figma can be done using AI agents
Writing code and code automation
AI agents can generate online music using previous music tunes and vector data
AI agents can perform Google Excel sheet data population, scraping, indexing, and analysis.
AI agent works locally within a computer and can even write the kind of articles which you are reading
Few AI agents use the browser scraping and crawling technique to perform browser-based or outside API tasks. For rest of the AI agent tasks such as creating a github repository, installing packages, generating directories and files, folders, writing basic code, and creating basic design on the Figma app can be performed using AI systems that integrate directly with APIs or leverage SDKs (Software Development Kits) to interact seamlessly with external applications. These AI agents rely on natural language processing and task-specific models to interpret user commands, automate workflows, and execute complex tasks efficiently. By combining generative AI capabilities with procedural logic, they can handle everything from setting up project environments to drafting initial prototypes, making them invaluable tools for developers, designers, and other professionals.
Resources for AI agents
We have github resources that provide github repositories of most of the AI agents.
Github repositories
https://github.com/NirDiamant/GenAI_Agents
https://github.com/SamurAIGPT/Best-AI-Agents
https://github.com/e2b-dev/awesome-ai-agents
https://github.com/Jenqyang/Awesome-AI-Agents
https://github.com/Yuan-ManX/ai-agent-roadmap
https://github.com/francedot/acu
Seem like a good story for today
See you in the next one
Summary
Nvidia's CES 2025 presentation detailed AI agent roadmaps.AI agents, non-robotic task performers, use logical reasoning and step-by-step processes. They leverage LLMs like ChatGPT, processing input, performing subtasks, and producing output iteratively. Langchain facilitates chain-of-thought prompting and context management. Multiple LLMs and frameworks support agent creation. AI agents automate tasks, boosting productivity.Nvidia's CES 2025 presentation detailed AI agent roadmaps.AI agents, non-robotic task performers, use logical reasoning and step-by-step processes. They leverage LLMs like ChatGPT, processing input, performing subtasks, and producing output iteratively. Langchain facilitates chain-of-thought prompting and context management. Multiple LLMs and frameworks support agent creation. AI agents automate tasks, boosting productivity.