> Agentic steps
What an agentic step is
When a request needs more than one move, Ocai breaks the job into steps. Each step is a unit of work with a beginning and an end: form a plan, take an action, check the result. The next step starts from what the last one found.
That is how Ocai works through a complicated request instead of guessing at an answer in one pass. Plan, act, verify, then plan again with what it now knows.
Thinking effort
Every model can spend more or less effort on each step. Thinking effort runs from zero through low and high to max.
- zero — fast and direct. Right for a clear question with an obvious route.
- low — some room to work things out.
- high — for problems with several parts, or where a wrong turn is expensive.
- max — the hardest work: multi-step analysis, awkward data, anything where a careless answer costs more than a careful one.
Because effort applies per step, what you spend tracks the difficulty of the job rather than a setting you had to guess at in advance.
What happens inside a step
A step can involve three things:
- Thinking — working out what to do next, at the effort level you chose.
- Tools — reading your files, running code, searching the web, calling a connected app.
- Checking — looking at what came back before deciding whether to continue.
Steps can also branch. A request that covers several independent pieces fans out into work that runs in parallel, then comes back together as one answer.
Asking for complex work
- Say what you want, and say what done looks like.
- Name the files or sources that matter, so the first step starts from real material.
- Raise thinking effort when a problem has several moving parts or a wrong answer is costly.
- Ask for autonomous mode when you want the whole chain run without you in the middle. See Modes.
Following along
- Read the steps as they run. Each one says what it is about to do.
- If a step is heading the wrong way, interrupt and steer. You do not have to let a run finish before changing your mind.
- Check the results against your own material before you accept them.
What a step costs
Steps are where credits go. Longer chains of steps cost more, and so does higher thinking effort inside a step. The Credits page breaks down how input, cache read, and output tokens add up.
Questions people ask
What happened to Deep Think, Deep Dive, and Deep Research?
They were separate modes once. Agentic execution is now built into every model, and thinking effort gives you the control those modes used to provide. Deep Think became a step where the model spends real effort. Deep Dive became more steps at higher effort. Deep Research became the plan, search, read, verify pattern run as a chain of steps.
Do I need a different model for hard problems?
No. Pick the model that fits the task, then raise thinking effort when the problem deserves it.
Can I see the steps?
Yes, they are shown as they run. That visibility is the point: you can see what Ocai is doing before it gets far.