intro to agents #92
Open
intro to agents #92
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
EricThomson
requested review from
Habeebah157 and
roshansuresh
and removed request for
roshansuresh
Habeebah157
reviewed
Jan 26, 2026
Collaborator
Habeebah157
left a comment
Habeebah157
left a comment
There was a problem hiding this comment.
Looks good, I dont see anything wrong
Collaborator
roshansuresh
left a comment
roshansuresh
left a comment
There was a problem hiding this comment.
This is a great intro lesson to agents! I've left a few comments/questions for clarification but this looks good to go. Also appreciate the addition of MCP like we had discussed!
| @@ -0,0 +1,45 @@ | |||
| # Introduction to AI Agents | |||
| AI agents are autonomous systems that can perform tasks on behalf of users by leveraging a combination of external tools, and decision-making processes. A helpful way to think about the difference between an agent and a traditional LLM is that a basic LLM answers questions, while an agent is able to perform actions using *tools*. This is why agentic AI has become so popular recently: it breaks LLMs out of the the "text-in, text-out" box, and let's them interact with the (software) world in much more interesting ways. | |||
Collaborator
There was a problem hiding this comment.
Might also be good to point out that agents are able to determine when they have completed a task, as opposed to a traditional LLM that stops once the response is generated
Collaborator
There was a problem hiding this comment.
You actually discuss this later. You can disregard this comment
|
|
||
| This pattern helps agents avoid brittle, one-shot responses. Instead of trying to solve everything at once, the agent can think step-by-step, verify intermediate results, and adjust its plan as needed. In theory, this makes agents more reliable and easier to debug, because you can see not just the final answer, but the sequence of thoughts and actions that led there. | ||
|
|
||
| ReAct is not a single library or tool: it is a design pattern that includes LLMs in the loop. Many agent frameworks employ some variation of this loop under the hood. |
Collaborator
There was a problem hiding this comment.
the phrase "includes LLMs in the loop" makes it feel like ReAct was developed independently of transformers and LLM theory. Is this true? If not, maybe we can reframe it: ReAct is not a single library or tool: it a response framework that leverages the inherent reasoning ability of LLMs.
| ReAct is not a single library or tool: it is a design pattern that includes LLMs in the loop. Many agent frameworks employ some variation of this loop under the hood. | ||
|
|
||
| ### Frameworks: smolagents, LangChain, and LlamaIndex | ||
| While building an agent from scratch is possible (indeed, we will build a couple in order to demysify agents and LLM tool use). However, just like with RAG, things can get complex very quickly, and there are [many agentic frameworks](https://github.com/Azure-Samples/python-ai-agent-frameworks-demos/) that have been created to handle this complexity for you. Just to name a few: |
|
|
||
| **[smolagents](https://huggingface.co/docs/smolagents/en/index)** is a lightweight framework from HuggingFace that emphasizes simplicity and transparency. It is especially well-suited for educational settings, because the agent loop is explicit and easy to inspect. Code-based agents are a first-class concept, which makes smolagents a good fit for learning how agents actually work under the hood. smolagents is the framework we will be using for the hands-on portion of this lesson, partly because it is simple and easy to learn, and because their code-based agents are so powerful and flexible. | ||
|
|
||
| **[LangChain](https://www.langchain.com/agents)** is a general-purpose framework for building LLM-powered applications, including agents. It provides abstractions for tools, memory, chains, and agents, making it easier to assemble complex systems quickly. The trade-off is complexity. LangChain can feel heavy, and understanding what is happening internally requires more effort. We initially planned to use LangChain for Code the Dream, but the learning curve was much too steep for our purposes. |
Collaborator
There was a problem hiding this comment.
what does chains mean in this context?
|
|
||
| **[LangChain](https://www.langchain.com/agents)** is a general-purpose framework for building LLM-powered applications, including agents. It provides abstractions for tools, memory, chains, and agents, making it easier to assemble complex systems quickly. The trade-off is complexity. LangChain can feel heavy, and understanding what is happening internally requires more effort. We initially planned to use LangChain for Code the Dream, but the learning curve was much too steep for our purposes. | ||
|
|
||
| **[LlamaIndex](https://developers.llamaindex.ai/python/framework/understanding/agent/)** isn't just for RAG, but also has an agentic framework that can be very powerful for building agents that interact with external data sources. |
Collaborator
There was a problem hiding this comment.
It seems like logically we should include a reason why we don't use this for our hands-on portion. But we don't really need to do that, we can discuss
| ### Tool-Based vs. Code-Based Agents | ||
| Not all agents work the same way. A useful distinction that is becoming more prevalent is between tool-based agents and code-based agents. | ||
|
|
||
| *Tool-based agents* interact with the world by calling predefined tools that are explicitly enumeratred to the LLM. These tools might include things like a plotting functions, a calculator, or any other well-defined software tools. The agent's job is to decide which tools to use and when, given the task. This approach is generally safer and easier to control, because the agent can only do what the available tools allow. |
Collaborator
There was a problem hiding this comment.
typo: a plotting function
|
|
||
| *Tool-based agents* interact with the world by calling predefined tools that are explicitly enumeratred to the LLM. These tools might include things like a plotting functions, a calculator, or any other well-defined software tools. The agent's job is to decide which tools to use and when, given the task. This approach is generally safer and easier to control, because the agent can only do what the available tools allow. | ||
|
|
||
| *Code-based agents*, on the other hand, are given free reign to generate and execute novel code to reach their given goal. Instead of selecting from a fixed set of tools, the agent writes code, runs it, inspects the output, and continues from there. This is extremely powerful and flexible. While careful sandboxing and guardrailes are required to keep things safe, in practice, it turns out that LLMs are surprisingly good at writing correct code, so code-based agents can be very effective. We will see an example of this in a hands-on lesson. |
| Also, the [Model Context Protocol](https://en.wikipedia.org/wiki/Model_Context_Protocol) (MCP), created by Anthropic, is an emerging standard that defines a *common interface* for exposing tools and resources to agents. It is a protocol that standardizes how agents discover tools, understand their inputs and outputs, and calls them safely, much like a USB standardizes how devices connect to computers. | ||
|
|
||
| ## Next steps | ||
| AI agents represent a shift from "text exchange" to "tool building"." The core ideas - autonomy, iterative reasoning, and tool use -- are simple, but powerful. In the rest of the lessons, we will move from these concepts to hands-on examples, where you will see how agents behave in practice. No newline at end of file |
Collaborator
There was a problem hiding this comment.
typo: an extra " added after "tool building"
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Brief introduction to agents: what are they, some different types, and a few of the frameworks out there. Just enough to whet their appetite for the lessons.