How Agents Work

The Agent Loop

A single API call to an LLM is not an agent. An agent is a loop:

1. Receive task
2. Think → what's the next step?
3. Act → call a tool (search, read file, run code, call API...)
4. Observe → get the tool result back
5. Repeat from 2 → until the task is done
6. Return final answer

This loop is what makes agents capable of multi-step tasks. A single LLM call can only work with what's already in the prompt. An agent can go get new information, take actions, and adapt based on results.

Tool Use

Tools are functions you define that the model can choose to call. You build them, the model decides when to use them.

Common examples:

search_web(query) - retrieve current information
read_file(path) - read a document
run_code(code) - execute and return output
send_email(to, body) - take an action in the world

Critical detail: The model doesn't execute tools itself. Here's exactly what happens:

1. You send the model a task + list of available tools
2. Model outputs: "I want to call search_web('AI agent frameworks')"
3. Your code actually runs the search
4. You feed the result back to the model
5. Model continues reasoning with that new information

The model is the brain. Your code is the hands.

Memory

Agents have three types of memory:

Type	What it is	Limit
In-context	Everything in the current conversation	Context window cap
External	A database the agent queries via a tool	Unlimited
In-weights	What the model learned during training	Fixed at training time

In-context is the default - everything the agent has seen this session. It's fast but finite.

External memory (via a vector database) is how agents access knowledge larger than the context window. This is what RAG is - covered in Phase 2.

In-weights is general world knowledge baked into the model. You can't update it without retraining.

Reasoning Patterns

Not all agents use the same thinking structure. Three common patterns:

ReAct (Reason + Act) - the most common

Thought: I need to find the current price of NVIDIA stock
Action: search_web("NVIDIA stock price today")
Observation: $875.23
Thought: Now I have the data, I can answer
Answer: NVIDIA is currently trading at $875.23

Plan then Execute - make a full plan first, then carry it out step by step. Good for complex, predictable tasks.

Reflection - after completing a task, the agent reviews its own output and checks for errors before returning the answer. Improves reliability on high-stakes tasks.

Where Things Go Wrong

Problem	What's happening
Infinite loop	Agent keeps calling tools without making progress
Tool misuse	Model calls the wrong tool or passes bad arguments
Context overflow	Long agent run fills the context window and breaks
Hallucinated tool calls	Model fabricates a tool result instead of actually calling it

These aren't edge cases - they're common. Handling them is a big part of what Phase 2 is about.

Module 0.3

The Agent Loop

Tool Use

Memory

Reasoning Patterns

Where Things Go Wrong

Sources