AI Agents Will Change How You Build Applications

AI Agents: A Deep Dive for Software Developers

Key Concepts:

• AI Agents: AI systems capable of advanced reasoning and interaction with external tools and the real world.
• Large Language Models (LLMs): Foundation models like ChatGPT and Gemini, often limited in reasoning and external access without agentic support.
• REACT (Reasoning + Act): A loop involving planning, reasoning, acting, observing, and repeating until a goal is achieved.
• Tools: Functions or code snippets that allow LLMs to interact with external systems (APIs, databases, files, etc.).
• Agentic Orchestration Frameworks: Tools like AWS Strands and LangGraph that facilitate building and managing AI agents.
• Dockstrings: Detailed descriptions of tools, crucial for LLM understanding and appropriate invocation.

---

1. Limitations of Traditional LLMs

Traditional Large Language Models (LLMs), like earlier versions of ChatGPT, exhibit key limitations when dealing with complex tasks. The speaker illustrates this with an example: a prompt requesting data retrieval, sorting, and filtering from a customer file. The LLM, without agentic support, produced a nearly correct result but failed to properly sort the data as requested, demonstrating limited multi-step reasoning capabilities.

A second major limitation is the lack of access to external data sources. While an LLM can process information provided within the prompt, it cannot independently retrieve data from databases, Google Drive, or APIs. This restricts its functionality to tasks requiring information beyond its pre-trained knowledge. The speaker emphasizes this extends to numerous real-world scenarios, such as retrieving current weather data.

2. The Rise of AI Agents & The REACT Framework

AI Agents address these limitations by providing LLMs with enhanced reasoning and access to external tools. The concept of AI Agents was popularized by the 2023 paper "Synergizing Reasoning and Acting in Language Models" (REACT). The paper highlights that prompting an AI to plan its approach to a problem significantly improves the quality of the solution.

The core of this approach is the REACT loop:

1. Reason: The AI analyzes the input and formulates a plan.
2. Act: The AI performs an action based on its plan (e.g., searching for information).
3. Observe: The AI receives the result of its action.
4. Repeat: The AI analyzes the result and iterates through the loop until the goal is achieved.

The speaker demonstrates this with an example of finding an author’s hometown: the AI reasons it needs a biography, searches for one, analyzes the results, and repeats until it finds the answer. Modern LLM applications like ChatGPT and Gemini now incorporate this agentic support, allowing users to observe the AI’s reasoning process in real-time.

3. Tools: Bridging the Gap to the External World

While agents provide reasoning capabilities, tools enable interaction with the external world. Tools are essentially pre-written code snippets that LLMs can invoke to perform specific tasks. The developer defines which tools are available to the agent using function definitions and detailed dockstrings – descriptions of the tool’s purpose, inputs, and outputs.

For example, an HTTP request tool allows the agent to retrieve data from websites. If prompted with "What's the weather in Toronto, Canada?", the agent would use the HTTP tool to query a weather API, parse the response, and return the weather information. Crucially, agents can only use tools explicitly granted access to, providing a mechanism for controlling the LLM’s behavior.

Two options exist for obtaining tools: utilizing pre-built tools from the AI community (for file system operations, shell interaction, web searches, etc.) or developing custom tools using agentic frameworks. Detailed dockstrings are vital for the LLM to understand when to use a particular tool.

4. Building Applications with Agentic Orchestration Frameworks

To build applications leveraging AI agents, developers utilize agentic orchestration frameworks. Two prominent options are:

• AWS Strands Agents: Developed by AWS, offering native integration with AWS services like Amazon Bedrock. It’s described as a higher-level framework, simplifying prototyping and experimentation. The developer provides the agent with tools and a goal, and the framework handles the rest.
• LangGraph: A lower-level, more feature-rich framework. Developers define nodes and edges to guide the LLM’s reasoning and can implement guardrails to prevent undesirable actions. It offers greater control but is more complex to implement.

The speaker recommends starting with Strands Agents for beginners before exploring LangGraph’s advanced capabilities. The example provided demonstrates a custom tool, get_customers, built using Strands Agents and annotated with @tool. This function takes a SQL statement as input and returns a list of records, with a detailed dockstring explaining its functionality.

5. Synthesis & Key Takeaways

AI Agents represent a significant advancement in LLM capabilities, overcoming limitations in reasoning and external access. The combination of agents and tools unlocks a new range of applications beyond simple chatbots, enabling complex problem-solving and real-world interactions. Agentic orchestration frameworks like AWS Strands and LangGraph provide the necessary infrastructure for building and managing these agents, offering varying levels of control and complexity. The speaker emphasizes the importance of detailed tool documentation (dockstrings) for effective agent behavior and control. The future of LLM applications lies in leveraging these agentic capabilities to create truly intelligent and autonomous systems.

Notable Quote:

“Tools are the secret sauce that enables LLMs to go much beyond chatbots.” – The speaker, highlighting the crucial role of tools in expanding LLM functionality.