Agent Factory

An agent is an LLM wired into a loop. It reads the situation, decides which tool to call, calls it, looks at the result, and goes again until the goal is met or a stop rule fires. That loop is what separates an agent from a chatbot, which gives one answer per message, a copilot, which suggests alongside you, and a fixed workflow, which runs steps you hardcoded.

The model chooses the steps, inside boundaries you set. It is the same shape as the harness from the class (a model, its tools, some context, and a loop), taken all the way to production.

The course breaks an agent into parts you can reason about: the model at the core, the tools it can call, the memory that persists across turns, the state the harness tracks, and the loop that ties them together. Get those right and the behaviour stays coherent. Get them wrong and it turns fragile or endless.

From there it moves through design patterns, planning, error handling and recovery, evaluation and testing, safety and guardrails, observability, and building with the major frameworks. Twenty-three modules and more than eighty wiki articles in all.

Most agent tutorials stop at a demo. This one aims at production: the parts that actually break, how to test them, and how to keep them safe.

If you understood the harness idea from the session and want to build the tools themselves rather than just use them, this is the deep version. It is by Akshat, one of the Oslo Vibe Coding organisers.