13 - Agent Foundations

An AI agent is a model that directs its own process: it reads its context, decides on an action (usually a tool call), observes the result, and repeats until the goal is met. This module builds that idea from first principles - the definition, the components, the loop, tools, memory, planning, where agents work and how to evaluate one - before later modules add protocols (13), multi-agent patterns (14), frameworks (15), production concerns (16) and harness engineering (17).

Learning objectives 9-11 hours (notes + lab)

By the end of this module you will be able to:

Decide whether a task needs a single call, a workflow or an agent, and at what autonomy level
Write a correct tool-use loop against the Claude and OpenAI APIs and against any OpenAI-compatible server, with bounds and error handling
Design tools, schemas and error messages a model uses reliably, and scale to large tool sets
Design working memory, task state and long-term memory (semantic, episodic, procedural) with their risks
Choose a planning strategy and know when reflection helps
Evaluate an agent on environment state with pass^k over repeated trials

Prerequisites

LLM Foundations - tokens, context windows, sampling
Prompt & Context Engineering - system prompts, structured outputs, context engineering
RAG - retrieval, which agents use as a tool

Where This Module Fits

flowchart LR
    W["01 What agents are"] --> AN["02 Anatomy"]
    AN --> L["03 The loop"]
    L --> T["04 Tools"]
    L --> M["05 Memory"]
    L --> P["06 Planning"]
    T & M & P --> PR["07 In practice"]
    PR --> E["08 Evaluation"]
    L -.-> LAB["🧪 Lab: loop from scratch"]
    E -.-> LAB

    style W fill:#e8e2d9,stroke:#ccc4b8
    style L fill:#e8e0d4,stroke:#c8b89a
    style E fill:#dde4dc,stroke:#b0c4b0
    style LAB fill:#d8dfe8,stroke:#b0bac8

Chapter Map

#	Chapter	You will learn	Time
1	What Are AI Agents	Agents vs workflows, the autonomy spectrum, why now, the decision test	40 min
2	Anatomy of an AI Agent	Model + harness: instructions, tools, memory, orchestration, guardrails, environment	35 min
3	The Agent Loop	Correct loops for Claude and OpenAI APIs, invariants, stop reasons, bounds, failures	60 min
4	Tool Use & Function Calling	Strict schemas, tool_choice, tool design, errors, idempotency, tool search	60 min
5	Agent Memory	One taxonomy; context management; checkpoints; Letta, Mem0, Zep, LangMem; poisoning	60 min
6	Planning & Reasoning	Reasoning models, ReAct, plan-and-execute, ReWOO, LLMCompiler, reflection, todo lists	50 min
7	Agents in Practice	Good use cases, personal vs enterprise, human-in-the-loop, build or buy	40 min
8	Agent Evaluation Basics	State-based grading, pass@k vs pass^k, first test sets, failure taxonomies	45 min
9	Q&A Review Bank	48 questions across the module	60 min

Code Lab

Lab	What you build	Runs on
Agent Loop from Scratch	A tool-use loop with validation and bounds over a small retail environment, graded on database state with pass^k across prompt and tool-description variants	Any OpenAI-compatible endpoint: a local Qwen3 on a laptop (mlx-lm, Ollama, vLLM) or a hosted API

Mini-Project

Build and evaluate a single agent for a domain you know (an internal help desk, a lab-inventory assistant, a personal finance helper):

An environment with 4-8 tools, at least two of them write tools with preconditions enforced in code, and a reset function.
A test set of 40 tasks across the six categories in Agent Evaluation Basics, each with an expected end state.
The loop from the lab (or a framework), with bounds, argument validation and informative errors.
Results: pass@1 with a confidence interval, pass^4, per-task counts, tokens and seconds per episode.
A failure taxonomy from the traces, one targeted fix, and before/after numbers.

Review

Q&A Review Bank - consolidated questions for this module
Module quiz - every Check Yourself question in this module, in course order

Previous: 12 - RAG · Next: 14 - MCP & A2A

Section Appendix

Summary & Key Terms - a quick recap of this section and its essential vocabulary.