Azure

Why Your AI Agent's Token Bill Exploded

Nothing failed. The loop just paid for its own history, twelve times per task.

What you'll learn

  • Why an agent loop’s input tokens grow with the square of its step count
  • What prompt caching actually charges — reads at 0.1×, writes at 1.25×
  • The one line that silently stops the cache being read, with no error
Azure 12 steps 51 seconds views

Was this useful?

Share

Understand it one step at a time

The short runs these in order in about 51 seconds. Here they are written out — pick any step to jump the short straight to it.

1 Step 1 of 12

The agent cost $4,260 last month

One customer-support triage agent. Four thousand completed tasks in a month. The invoice landed at six times the original estimate.

Frequently asked questions

Why does an AI agent's token bill grow faster than the number of steps it takes?
Most agent loops resend the entire conversation history with every step — step 10 re-sends everything from steps 1 through 9. That makes total input tokens grow roughly with the square of the step count, not linearly with it.
How does prompt caching change the cost of that resent history?
A cached read is billed at roughly 0.1× the normal input rate, while writing something new into the cache costs about 1.25×. The saving only materializes if the prefix you're resending is byte-identical each time — one line that changes something early in the prompt can silently break the cache and put you back to paying full price.

Free app · no app store

These are built for a phone

Every short is drawn at full portrait height, the shape a phone already is. Installed, it opens full-bleed with no address bar across the top — and the whole library reads offline.

How it works

Read deeper on Azure

Twenty seconds gets the shape of an idea across. These go into how it behaves in production.

More shorts

Get new posts by email

New technical articles, Azure AI and GitHub Copilot updates, and upcoming events. No spam, unsubscribe anytime.

navigate open