Stage 3 · Design · Intermediate
Agents Fundamentals
Build an agent from the loop up and learn when not to. Tell workflows from agents, write the tool-use loop with hard limits, design safe and teachable tools, manage planning and context, enforce safety in code, and evaluate agents with repeatable tasks and honest statistics.
7 lessons · 15 hr · Free
You will learn to
- ▸Choose between a single call, a workflow and an agent, and build the five standard workflow patterns
- ▸Write an agent loop with turn, token and error handling, and test it with a scripted client
- ▸Design sandboxed, bounded and self-explaining tools and test them like security boundaries
- ▸Manage planning, notes and long runs without editing conversation history
- ▸Enforce permissions, approvals, limits and loop detection in code, and defend against injected instructions
- ▸Evaluate agents on checkable tasks, repeat runs, report intervals and read traces to find failure causes
Before you start
- ·Comfortable with Python and the Claude API, as taught in Build with LLM APIs
- ·Recommended: Evaluate and Use AI Responsibly, for evaluation statistics and safety thinking
- ·API credit is only needed to run agents against a real model. All tests in the course run without one
Syllabus
- 01Workflows versus agentsThe definitions, five workflow patterns in code and when an agent is worth it.1 hr 30 min
- 02The agent loopBuild the loop, add limits and a trace, and test it without a model.1 hr 30 min
- 03Designing toolsSandboxed file tools, descriptions as prompts and errors that teach.2 hr
- 04Planning, memory and long tasksPlans, notes, small outputs and restarting from a summary.1 hr 30 min
- 05Safety and controlA policy layer, approvals, sandboxes, limits and injected instructions.2 hr
- 06Evaluating agentsEnd-state checks, repeated runs, pass-to-the-k and reading transcripts.1 hr 30 min
- 07Capstone: an incident investigatorBuild, test and evaluate a read-only log-investigation agent.5 hr