The Brain
Start here. What an agent is, in one sentence, and the nine pages in the order to read them.
An agent is a loop: the model reads the context and returns exactly one of two things, a tool request or a final answer. The runtime does everything else: it wakes the loop, builds the context, executes the tools, checks the budget, validates the output, and writes the result. The model never touches anything. It only asks.
Anyone in the house who works with the agents and wants to know what is actually happening when one runs: which part is the model, which part is our code, where the money goes, and where a human can step in. Written for a rookie, checked against the code.
Start with the words, then walk the loop. The rest are maps of the same loop from different angles. Every page stands on its own, and every page uses the same names. Names follow the Agent Circuit in the repo where the repo has a name, and the industry's usual name where it does not.
- 01The BrainStart here. What an agent is, in one sentence, and the nine pages in the order to read them.
- 02The WordsEvery word the other pages lean on, defined once, in plain English, with what other people call it.
- 03The Eleven StepsOne run, from a cold trigger to a committed result, and the flow chart it draws.
- 04Canon NamesWhat each step is called in the repo, in the API, and everywhere else.
- 05MisconceptionsThe things rookies believe about agents that are not true. Yours first.
- 06The Decision TreeFive places a run can fork, and who decides each one.
- 07One Full RunOddz on a game night, drawn two ways: the packet rule and the tool loop.
- 08Loop Versus AgentThe loop is the mechanism. The agent is the thing that owns it and survives between runs.
- 09The VariationsEight shapes you will meet, from no loop at all to loops made of loops.