Learning map / Task 1 / Lesson
Where the round trips come from
Five minutes. The page below runs the same scripted model the tests use: a program, not a language model, that follows the documented rules of tool use, so the same settings always give the same turns. Nothing is plotted from a formula; every turn you see was produced by running the loop with your settings.
The desk asks for the open orders of some customers and the status of each order. The model can only act by asking your code to run a tool, a tool call, and it only learns the result on its next request. One request and its reply is a round trip. Each request re-reads everything so far, counted in tokens, the pieces of text a request is measured and billed in . A tool can take one item, or a whole list: a batch tool.
The turns
Each row is one request. The gray bar is the size of what the model re-reads, in input tokens: the request's text, characters divided by four, the same estimate the task page uses. The chips under a bar are the tool calls the reply asks for; the last row is the reply that answers. The address bar carries your settings, so the link opens the same view.
All values at once
Round trips against the number of customers, for each tool shape, under the loop settings you chose above. Flip "how results go back" and watch which lines move, and which do not.
Work out, do not look up
- Find a setting where batch and composite cost the same number of round trips and different tokens. What is the token difference paying for?
- With one message per result, the first stage costs nothing extra. Where does the cost land, and why there?
- At what number of customers does the per-item shape stop being a toy problem for a human reviewer reading the transcript?
When you can answer these from the picture, open the task page and predict the four rungs, versions v0 to v3 of the agent, each changing one thing.