I built a set of small hands-on exercises on how to build and debug LLM agents with various capabilities. The suite runs locally and covers tasks like:

- general llm-style chat: math Q&A

- tool use: hashing with tools

- image understanding

- web-browsing: (tic-tac-toe)

- code generation & execution

- memory across sessions

The idea isn’t benchmarking, but giving builders practical drills to see where agents succeed or fail. Useful for education, workshops, or self-study.

Would love feedback on what other exercises would make it more useful.

Try it here: https://ape.llm.phd/