S24 · decision value simulation

Month-end commentary, a ten-step loop against one call

People, data and outputs in this scenario are simulated. Each stage is recorded when you commit; the reveal comes after.

4stages
35 minabout
0stages committed

The brief

A retail bank (simulated) writes a short commentary on each account manager's monthly portfolio variance report: 40,000 reports a month. The commentary must quote figures from the report correctly; a person reviews a sample.

Prices and figures are hypothetical. Model price: 3.00 USD per million input tokens and 15.00 per million output tokens, no caching.

Workflow W: code computes the variance table and top drivers; one model call (3,000 input tokens, 400 output tokens) writes the commentary; a validator checks that every figure in the text matches the table, with one retry.
Agent A: a loop with tools to query the report data. The first call has a 1,500 token prompt; every step re-sends the whole history and adds 700 tokens (tool results and reasoning); every step writes 250 output tokens.

Stage 1 of 4

For an n-step run the input tokens are n x 1,500 + 700 x n(n - 1) / 2, because each step re-reads everything before it. Output is n x 250 tokens.

What is the cost of one 10-step agent run?

0.0825 USD: ten calls at 1,500 input and 250 output tokens each, since the prompt is fixed and the loop simply repeats it.0.1395 USD, because output tokens are a small part of the work and the input side is what grows with the history.0.2376 USD, the cost of a 12-step run, since the loop has a cap of 12 and should be budgeted at its cap.0.177 USD: 46,500 input tokens (15,000 + 31,500) at 3.00 per million is 0.1395, plus 2,500 output tokens at 15.00 per million is 0.0375.