SUN, SEPTEMBER 27, 2026
Independent · In‑Depth · Practitioner‑Tested
Claude Productivity

Seven Prompts for Agent Unit Economics That Survive Scale

Cognition is valued at $48 billion on roughly $900 million of revenue - about 53x, the same multiple its May round paid. Nobody has published the gross margin, because autonomous agents burn tokens proportional to how hard the problem turns out to be, not to how much output they produce. These seven prompts build the cost model your own agent deployment needs before the invoice explains it to you.

⌨️ 7 prompts 🕐 Updated Sep 27, 2026
💡 How to use these prompts: Replace everything in [BRACKETS] with your specific details before sending. Click Copy to copy any prompt to your clipboard instantly.
1
Cost Per Merged Pull Request
An agent at half the price that lands half as many PRs is more expensive. This is the only comparison that settles it.
Here are my agent logs and the outcome of each run.<br/><br/>Do not give me cost per million tokens. Give me cost per pull request that was actually MERGED.<br/><br/>Break out: total spend, runs that produced a mergeable PR, runs that produced a PR that was rejected or rewritten, and runs that produced nothing.<br/><br/>Then tell me what share of total spend went to the last two categories. That is the number I want.<br/><br/>LOGS:<br/>[paste]
2
Does Cost Scale With Difficulty or Output?
Autonomous agents burn tokens proportional to struggle, not to lines produced. Most budgets assume the opposite.
Below are my agent runs with the size of the change produced and the tokens consumed.<br/><br/>Test whether token spend correlates with output size or with something else - number of files touched, test failures encountered, retries, time spent.<br/><br/>Tell me which variable actually predicts cost, and give me a formula I can use to estimate a task's cost before running it.<br/><br/>If the data does not support a clean predictor, say so rather than fitting one.<br/><br/>RUNS:<br/>[paste]
3
Find the Tasks That Never Pay Off
Every agent deployment has a set of tasks it quietly loses money on. Find them before the finance review does.
From the task history below, identify categories where the agent's success rate is low enough that supervised work would have been cheaper.<br/><br/>For each category give: attempt count, success rate, average cost per attempt, average cost per success, and an estimate of what a developer would have cost.<br/><br/>Recommend which categories to stop sending to the agent entirely. Be specific - I would rather cut three categories than shave 5% off everything.<br/><br/>HISTORY:<br/>[paste]
4
Add the Monitoring Overhead
Monitoring is not optional at scale and it is rarely in the original business case.
My current agent inference cost is below. OpenAI has disclosed that continuous behavioural monitoring costs about 20% additional inference compute.<br/><br/>Recalculate my cost per completed task with that overhead applied, and show the break-even against a human doing the same work.<br/><br/>Then tell me at what success rate the agent stops being cheaper than the human, with monitoring included.<br/><br/>COSTS:<br/>[paste]
5
Model the Price Rise
Cheap inference is an acquisition strategy. Plan against the price you will pay in year two.
My agent runs on the model below at current list prices.<br/><br/>Model three scenarios: prices rise 2x, 3x, and 4.5x - the range DeepSeek actually moved this year.<br/><br/>For each: my new monthly cost, whether the deployment is still cheaper than the human alternative, and what I would have to change to stay viable.<br/><br/>Then tell me which of those changes I should make now regardless, because they are good ideas at current prices too.<br/><br/>USAGE:<br/>[paste]
6
Separate Review Cost From Agent Cost
The hidden cost of unattended agents is reviewing work you did not watch being made.
For each completed agent task below, I have recorded how long a human spent reviewing the output.<br/><br/>Calculate the fully-loaded cost per task including review time at my engineering hourly rate, and compare it against the cost of that engineer doing the task themselves.<br/><br/>Flag any category where review time is so high that the agent is producing negative value.<br/><br/>Asynchronous agents shift work from doing to reviewing - I want to know where that trade stopped being worth it.<br/><br/>TASKS AND REVIEW TIMES:<br/>[paste]
7
Write the Number for the Finance Review
Writing this before you are asked for it is how the deployment survives the first budget cut.
Using everything above, write a one-page summary of my agent deployment's economics for someone who does not work in engineering.<br/><br/>Include: total spend, tasks completed, cost per completed task, the human-equivalent cost, net position, and the three largest sources of waste.<br/><br/>Rules: no claim without a number behind it, state uncertainty where the data is thin, and do not round in my favour.<br/><br/>If the deployment is not paying for itself, lead with that.<br/><br/>DATA:<br/>[paste]