AI Agents¶

AI agents are simply workflows, or DAGs, or just graphs if included on a loop. Most steps in these workflows should be regular code, not LLM calls.

Fundamental Building Blocks¶

1. Intelligence¶

This is the only truly AI component. Without it you just have regular software.

flowchart LR
user["`User input`"]
llm["`LLM Processing`"]
response["`Generated response`"]


user --> llm --> response

Cookbook - LLM Completion

2. Memory¶

LLMs don't remember anything from previous messages. Without memory, each interaction starts from scratch because LLMs are stateless.

3. Tools¶

The LLM call itself is limited just to text, but you want to do more than that, such as API calls and interact with databases.

Cookbook - LLM Tools

4. Validation¶

LLMs are probabilistic and can produce inconsistent outputs, make sure that the LLM return structured output A.K.A JSON format that matches your expected schema. If does not you can send it back to the LLM to fix it. Most commonly used tools are Pydantic (Python) and Zod: (Javascript).

5. Control¶

You don't want your LLM making every decision. Use deterministic control flows like if/else statements and routing logic. This is normal business logic written in any kind of application.

Some times preferred instead of tools why it can be a little difficult to understand why the LLM have chosen that tool. With routing based control flows and structured outputs we can define reasoning output properties so it is easier to debug.

Cookbook - Control Flow

6. Recovery¶

Things may go wrong. APIs will be down, LLMs will hallucinate and rate limits will hit you. You need try/catch blocks, retry logic with backoff, and fallback responses when stuff breaks. This is just standard error handling that you'd implement in any production system.

7. Feedback¶

HITL (Human in the loop)

Human approval because some decisions are too important or complex for full automation. Like sending e-mails or making purchases. Add approval steps where humans can review and approve/reject before execution. This is just a basic approval workflow like you'd build for any app.

AI Agents vs. Workflows¶

Feature	🟢 Agent	🔴 Workflow
How many LLM calls?	Many	Many
Who decides when to stop?	The LLM	Predetermined Steps
Good for?	When the path to the solution is unclear	When the path to the solution is known upfront