How I Built My Own AI Agents From Scratch
Session Description
In this talk, we’ll learn to build our own custom AI Agent.
We’ll start by answering a few basic but important questions: What is an agent? How does it “think”? What do we mean by tools, models, and memory in this context? From there, we’ll go hands-on.
In this demo will use Agno, a full-stack open-source framework for building AI agents with tools, memory, and reasoning, and Streamlit for a minimal user interface. I’ll walk through how I created a custom agents using Agno:
Agents that are not just chatbots but also for scraping, analysis, automation, and domain-specific assistants that use external knowledge bases, depending on your needs.
How to choose a model
How to add tools (e.g. stock data, search APIs)
How to give the agent clear instructions
How to use memory or build a team of agents when needed
I’ll show live how everything connects together in the Streamlit app, from tool setup to the agent responding intelligently to real-world queries.
We’ll also discuss what can break (like bad tool responses, unclear goals), how to debug those issues, and what kinds of tasks agents are actually useful for.
This is not a high-level overview, it’s a grounded, hands-on session for anyone curious about building functional agents using opensource tools. If you've seen demos of “AI assistants” and wondered how they’re built, this talk will show you exactly that, with real code and zero marketing. You don’t need any prior experience with AI or ML, just a basic understanding of Python is enough to get started.
Key Takeaways
What it means to “build an AI agent”
How to define tools, tasks, and goals for agents
A working example you can build and extend on your own
Common mistakes people make when building agents, and how to avoid them
How to stay in control of what your agent does
Session Categories
Which track are you applying for?
Speakers
Shivam Katare
Reviews
On reading this description, it makes more sense to be on workshop rather than talk.
Makes more sense as a workshop, you will not be able to cover anything substantial in a talk slot.
The proposal is thorough, but tool-use is the focus of the talk, and the topic isn't novel. Although the tool in focus is a FOSS project, this doesn't warrant a main track talk, in my humble opinion.
There seem to be two parts to the proposal content:
getting an AI agent system up & running
making it better
I think the content can be better presented as:
a quickstart document or a lightning talk - for #1
a workshop - to cover #2 well
Reviewers felt that the topic of using tools with AI agents is not novel enough to warrant a main-track talk, even though the project you mentioned is open source.
While they appreciated its thoroughness, the consensus was that the content would have been better suited for a workshop rather than a standard talk or a lightning talk, as there's a lot of content to cover.