Agentswhen words become actions

0:03Text that counts as commands

Everything a model emits is text. But once we agree that certain strings count as instructions, emitting text becomes doing things. One line is ordinary thought: check what videos exist. The next line is a command, waiting for something to execute it.

0:22The agent loop

That agreement produces a loop. The model thinks and emits a tool call. The environment — everything outside the model — executes it. Whatever comes back gets added to the end of the context, as text the model can read.

0:37Think, act, observe

It reads what came back and thinks again. Think, act, observe — lap after lap, until the task is done, or the model decides it is.

0:48Tools as text interfaces

A tool is anything with a text interface. Search the web, run code, edit files, drive a browser. If it takes text in and hands text back, it plugs into the loop.

1:02Context as working memory

The context window is now the agent's working memory. The task, the thoughts, the calls, the results — laid end to end in one sequence. The machinery underneath is unchanged: the Transformer reads everything so far, predicts one next token, and that token feeds back into the context. Agency, built out of autocomplete.

1:26This series as proof

This series is proof. A language model sat in this loop, with a terminal and a renderer for tools, and wrote, rendered, and revised every one of these lessons. The words a model emits became commands, the commands became actions, the actions became these pictures. Agency was autocomplete all along.