🎯 Determinism

New here? Determinism means doing the same thing, the same way, every time—without “deciding” anything along the way. It’s the difference between a cake recipe (always the same steps, in the same order) and an experienced cook (who tastes, adjusts the seasoning, and changes plans if an ingredient is missing). Both can make a good cake, but one is predictable and the other is flexible.

In the previous modules in this track, you saw the agent “thinking”: it reads the context, decides which tool to use, evaluates the result, and adjusts the next step — that’s how the routine that learns (module 6.4) improves on its own over time. Determinism is the opposite of that, by design: a fixed sequence of actions that doesn’t change, interpret, or “think” — it just executes.

Caption: on the left, the “thinking” agent goes through a decision cycle at every step (useful when the task varies). On the right, the deterministic script follows a straight, fixed line of numbered steps—with no cycle and no decisions along the way.

It may seem strange to ask an agent for less "intelligence," but it makes perfect sense for a specific type of task: repetitive and well-defined. If every day at 7 a.m. you need to export a report, rename the previous day's files, and email it — the task's format doesn't change. It doesn't need "judgment"; it needs reliability and speed.

An agent that “thinks” every time uses more tokens (the unit of processed text that determines cost, covered in module 0.1) and takes longer, because it reevaluates the context from scratch on each run. A fixed script skips that step: it already knows exactly what to do.

New here? One script is a fixed list of instructions the computer runs in order, one after another, without interpreting or deciding anything— it’s the recipe from the previous example, only written in code instead of Portuguese. Unlike the agent, which “reads” the situation and chooses what to do, the script simply follows exactly what’s written, the same way every time.

The good news: the person writing the script can be the agent itself. It can "think" once, put together the right sequence of steps, and save it to a file. From then on, whoever runs it doesn’t need to think — they just run the file.

Here’s the trick that separates an expensive scheduled task from a cheap one: instead of asking the scheduled task to “think” every time it runs, you ask the agent write the script just once — and from then on, the scheduled task only runs that script. The "intelligent" part (writing, reviewing, adjusting the script) happens rarely and separately; the "repetitive" part (running it every day) is fast and inexpensive.

Caption: on the left, the agent thinks through and writes the script ONCE. On the right, the scheduled task runs that ready-made script EVERY day—fast, without spending tokens to “think” again on each run.

Objective: ask the agent to write a simple script for one of your repetitive tasks and configure a scheduled task to run only that script from now on — without "thinking" again each time it runs.

Quero transformar a tarefa "<descreva a tarefa repetitiva, ex: exportar o relatório de vendas
do dia anterior e salvar em /relatorios>" num script determinístico.

1. Escreva um script simples (bash ou python, o que fizer mais sentido) que execute exatamente
   essa tarefa, sempre da mesma forma.
2. Teste o script uma vez rodando manualmente e me mostre o resultado.
3. Configure uma tarefa agendada para rodar ESSE SCRIPT (não o agente pensando) no horário
   "<defina o horário, ex: todo dia às 7h>".
4. Explique em uma frase por que isso fica mais rápido e barato do que o agente decidir tudo
   de novo a cada execução.

How to verify: run the scheduled task manually once and check that it calls the script (it doesn’t open a new "thinking" session); the runtime should be much shorter than a normal agent session.

Determinism is great until your task needs to evaluate a changing context. If the right decision today differs from the right decision tomorrow—because it depends on something that varies (email volume, document content, an external condition)—a fixed script will get it wrong because it can't adapt. In that case, having the agent “think” (covered in module 6.4) is worthwhile again.