# When monitoring stops scaling

Source: https://dinand.com/research/when-monitoring-stops-scaling/
Author: Dinand Tinholt

> What a simulation suggests about the cost of overseeing a growing agent portfolio.

In the AGENESIS-2 simulation, modeled portfolio cost grew by about 2.76 times each time the agent portfolio doubled while the monitoring team stayed the same size. That corresponds to a scaling exponent of 1.466. A design that gave each agent its own check grew roughly in proportion to the portfolio.

## What the result puts into focus

The useful question concerns the capacity of the oversight design. If agents expand faster than the ability to check their work, adding agents changes the operating conditions for the whole portfolio. A budget based on cost per agent can miss that interaction.

A CIO can use this as a reason to test several portfolio sizes before committing to a rollout. Measure the work required to review decisions alongside the consequences of missed errors. Examine whether the review process still works under the larger workload.

## Correction and limits

The supplied executive brief describes 256 pre-registered simulation units. Independent verification identified 95 duplicate runs that differed only in a setting with no effect at most portfolio sizes. Correcting the arithmetic left the headline exponent at 1.466; one reported verdict changed.

These are simulation findings from the supplied study summary. This page does not establish an empirical scaling law for deployed enterprise agents. The raw ledger and full methodology are not included here. The accompanying explorer simply illustrates the reported exponent; it does not forecast actual costs.

Continue the thread

## A next room to enter.

[Choose another path ](https://dinand.com/)

[Research · Published article overview

### When liability changes the reviewer

A controlled simulation of how a record of past harm affected an AI supervisor.

](https://dinand.com/research/oversight-and-liability/)[Research · Paper

### When watching every hour starts to pay

A pre-registered simulation held hourly monitoring of an agent fleet and six-hourly inspection to the same nominal reviewer budget, and found the budget at which hourly monitoring starts to catch faults.

](https://dinand.com/research/when-watching-every-hour-starts-to-pay/)[Research · Study summary

### The Judge’s Mirror

How often a model judge agrees with itself on 197 blind case files, and how far down one model family a judge stays usable.

](https://dinand.com/research/the-judges-mirror/)

[All Research ](https://dinand.com/research/)[Back home ](https://dinand.com/)

---
Site guide for AI agents: https://dinand.com/llms.txt