Patrik Reizinger gave this talk at the 50th Machine Learning Summer School on 8 September 2026, a two-hour session on what agents are, what research is for, and how to run agents inside research. This page is the principles from that talk, rebuilt to be read rather than presented: each principle is its key points, the talk's own drawing where it had one, a line for practice, and an example.
Agent, on this page, means the same language model as in a chat window, given files, a computer and permission to act on your behalf; section 3 says what each addition changes. Ten minutes? Read the bold line under each heading, then do the audit in section 5. Where a tool appears (a hook, a rule, a skill, a test), it is an example of a principle, folded under its title so you can open it and copy it.
Five parts: why do research at all, how taste is trained, what an agent is, how to run one as a supervisor, and an audit for the next task you think of delegating.
Your answers and read marks stay in this browser; nothing is sent anywhere.
Posts are from The Path to PhD, Patrik Reizinger's newsletter, unless noted; “P2P No. N” is its issue number.