One project, 70 AI helpers: why AI should audit how your agents work
Own Your Systems · November 20, 2026 · 4 min read

One project, 70 AI helpers: why AI should audit how your agents work

A case study from LaPage's own AI agents: one website project called 70 helpers and two agents used almost two thirds of a day's AI cost. The four changes we made, and the lesson for businesses bringing AI into operations.

LP
LaPage Digital
Systems Editorial
One project, 70 AI helpers: why AI should audit how your agents work

At LaPage, most of the daily work is done by AI agents. Each area has its own agent: one looks after websites, one builds internal software, one handles sales, one writes content, and one coordinates the rest. When a job is big, an agent may call in "helpers" (more AI) to split the work and finish faster.

It sounds tidy. Until the day we asked the AI to check how much the whole system had spent in a single day.

One project, 70 helpers

AI is billed in tokens, roughly the amount of text the AI reads and writes. The one-day summary showed:

Two agents used almost two thirds of the day's AI cost

  • Website agent: 34%. For a single website project it had called 70 helpers. Each helper ran 100 to 180 steps, mostly on the most expensive model.
  • Internal software agent: 32%, also with more than 70 helpers.
  • The dozen-plus other agents combined: about a third.

Two agents took almost two thirds of the day's cost. The work still got done and the results were usable. But it cost several times more than it needed to.

The numbers showed one more leak: a few other agents were carrying far too much in their heads. The longer a working session runs, the more old conversation the AI has to re-read at every step, so each step costs more than the last.

Nobody meant for this to happen

There was no bug here. An agent given a job simply gets it done. A big job? Split it up and call in more hands. Not sure? Use the best model to be safe. It never asks itself "is this too expensive?".

It's like a capable, hard-working employee nobody has told the budget to. That person isn't wrong. Setting the limits is the manager's job.

What we changed right away

Four rules so AI works well and cheaply

1. One small, clear task per helper. Don't hand over a whole lump and let the AI split it. The clearer the task, the fewer steps a helper takes and the less it has to guess.

2. Cheaper AI for simple work. Translating, proofreading, checking images and data entry don't need the top model. Save the expensive one for design, hard logic and judgment calls.

3. Summarize long sessions, then continue. When a session runs too long, the agent writes down what it needs to remember in the job file and starts again, lean, from there.

4. Most important: AI checks how the agents work, every day. A clear table: how much each agent used, how many helpers it called, which models it ran. No waiting for the monthly bill to get a surprise.

That table is exactly what showed us the problem. Without it, everything would still have been "working fine", with only the bill quietly growing.

The lesson for businesses bringing AI into operations

AI can do a lot of work. But AI does not replace management. Bringing AI into a company is like hiring: you still need to know

  • what it is working on,
  • what each piece of work costs,
  • and who reviews it regularly.

The simplest way is to have the AI do that check itself, on a schedule, with numbers. One table a day, a few lines of notes, and one person who reads it.

There is one condition: the system has to be yours. When you run AI on tools you own, you can read every number and set your own rules. Rent a black box by the month and all you see is the total at the end.

How LaPage can help

LaPage builds AI and automation systems that the business owns, with cost and behaviour tracking built in from day one, not bolted on after the bill has grown. If your team uses AI and nobody is tracking what it costs, send us a message and we'll look at it together.

And on your side: what are you using AI for, and who is keeping an eye on what it costs?

LP

LaPage Digital

LaPage Digital publishes systems-focused writing on architecture, automation, infrastructure, and operational delivery.

Editorial subscription

Stay Updated

Subscribe to our newsletter to receive technical insights and updates on best practices for web development and hosting infrastructure.

We respect your privacy. Unsubscribe at any time.

Call: +84 981 690 658
Message us on Facebook
Zalo
Contact us on Zalo
Message us on WhatsApp
Contact us