Lucek traces the progression from text completion to instruction tuning, function calling and ReAct-style agent loops. In his account, a useful agent does not simply generate an answer: it plans a step, invokes a tool, observes the result and updates its next action.
A browser-use demonstration shows an agent searching for Lucek's website, extracting page structure and gathering information for a summary. The browser becomes a controlled tool surface while the model receives observations from each navigation and interaction step.
He then demonstrates a desktop agent inside an isolated Ubuntu environment. The model works from screenshots and GUI actions to operate LibreOffice, illustrating how computer-use agents can handle applications that do not expose convenient APIs.
A final ChatGPT agent example researches a topic and builds a presentation over a longer run. Lucek argues that these systems are most useful when their tools mirror the digital environments people already use, while secure isolation and visible intermediate steps keep the work reviewable.
Watch on YouTube
