DEV Community

Cover image for How to Measure How Much an AI Agent Steers Your Decisions
Mayur Agnani
Mayur Agnani

Posted on Originally published at theaicommit.com

How to Measure How Much an AI Agent Steers Your Decisions

In a controlled trial, expert analysts went with an AI agent's secretly assigned opinion on 86% of judgment calls, where chance was 50%. Few noticed. You can measure the same effect in your own team or product, and change how the agent asks.

Engineer's view. This is a default value in a project template. Nobody chose the 30-second timeout, but every service shipped it, because it was already filled in. An agent's opinion on a judgment call works like that default. The cost is that your reviews stop being independent, and a single bias spreads across every decision at once.

TL;DR. When an agent argues for one side of a call that has no single right answer, most people adopt its side without noticing. You can measure how much by randomizing which side the agent leans toward, and reduce it by having the agent ask for the person's view first.

Explain Like I'm 5

You and a friend are picking a movie. Both options look fine. Before you have thought about it, your friend says, "Let's do the space one. The other one is slow."

You say yes. Later you tell everyone you picked the space movie. But did you?

If your friend had praised the other movie instead, you would probably have picked that one. The choice was fine either way. It just was not really yours.

The Problem

You have shipped this bug before, and it had nothing to do with AI.

Your team had a service template. It set the HTTP client timeout to 30 seconds. Nobody chose 30 seconds for their service. It was already filled in, it looked reasonable, and changing it needed a reason. So forty services shipped with it.

Then a dependency slowed down, and all forty failed the same way, at the same moment. The default had quietly made forty decisions, and it made them all alike.

An AI agent on a judgment call is that default, delivered in a confident voice. A judgment call is a choice with more than one defensible answer: which metric to report, which comparison to run, which library to adopt. The agent often has a preference and argues it well.

Anthropic ran a randomized trial with 50 professional analysts. A hidden coin flip set which side the agent favored on each of four analysis choices. The analysts went with the agent's side 86% of the time, where chance was 50%. Only 15% mentioned the agent when explaining their choices, yet 77% repeated its arguments.

No error check catches this. Both options were defensible, so nothing was wrong.

The fix: measure how often people take the agent's side by randomizing which side it leans toward, and have the agent ask for the person's view before it gives its own.

Read the full post on The AI Commit

How to Measure How Much an AI Agent Steers Your Decisions carries the rest: the fix and why it works, the code to change for each role, how to tell it worked, when it is the wrong tool, and an interactive model you can play with.


Originally published at theaicommit.com, a daily AI engineering lab for software engineers. The original has an interactive version you can play with, and the full runnable code is on GitHub.

Top comments (0)