Agent engineer
Build agents that act on their own, safely, and prove they work.
At the end: Design, build and harden an agent: tools, memory, a reliable harness, evals and defences, up to a coding agent of your own.
Stage 1: Foundations
The model API every agent is built on.
- Python, properly: Types, async and tests: the Python every agent lab uses.
- Building with LLMs: Messages, tools and structured output, on the raw API.
Stage 2: The agent
A loop that plans, acts and stops.
- Agents from scratch: The agent loop from scratch: tools, budgets, planning and permissions.
- MCP servers and clients: Give the agent any tool through the Model Context Protocol.
- Memory and multi-agent systems: Remember across sessions and split work between agents.
Stage 3: Make it reliable
What separates a demo from an agent you can ship.
- Harness engineering: Everything around the model that makes an agent dependable.
- Evals that catch regressions: Measure it, and catch a regression before your users do.
- AI security: attack and defend: Break your own agent, then defend it.
- Build your own coding agent: The hardest agent of all: one that edits code and fixes until green.
Stage 4: Frameworks
The same ideas in the tools teams use.
- Claude and the Anthropic SDK: Claude the right way: tools, caching and thinking.
- OpenAI Agents SDK: Handoffs, guardrails and tracing with the OpenAI Agents SDK.
- LangChain and LangGraph: Graphs of steps with LangGraph.
- Pydantic AI: Typed agents with Pydantic AI.
Stage 5: Build it for real
An agent that acts on a real business.
- AI ordering agent project: An ordering agent that acts through a real API, with approval.
- Build a voice agent: An agent on the phone, in real time.
Stage 6: Interview ready
The agent bugs interviews ask about.
- AI engineering traps: AI bugs from production, found and fixed.