Delegating decisions to AI

On one hand it sounds utopian, but listen to what the main idea is.

If we are creating an autonomous system, or want it to do most of its work autonomously, we need to give it enough rules, protocols, conditions, limits and other things that will guide the agent. Because an agent can very easily drift into completely inefficient loops, doing things that do not need to be done at all, and so on.

I have been coding with agents almost nonstop for a year and a half, and this is a very typical problem. It creates at least two difficulties.

It can look like the agent is doing something useful, while in fact it has already wandered off somewhere into nowhere. This spends resources and creates additional, completely unclear solutions that you may never have asked for.

And if it brings every decision point to you, you have to keep sitting with the agent and waiting for the next question. Roughly speaking, you only need to make a decision from time to time, but to do that you have to stay nearby all the time. Also inefficient.

I try to systematize these things as much as possible, especially recurring tasks. Collect everything possible about a task so that the agent already has answers for different cases and situations.

Here is the mechanism I am introducing into my own practice. In the next cycle we will start introducing it into the product with the team.

I asked an agent to go through all my threads from the past year and find every decision point I had resolved and exactly how I had resolved it. A question, a sticking point, a problem, and my feedback: what I told the agent, what decision I made.

The task is to pull this out of all the different areas where I do development and understand what types of decisions I make. Which ones are more predictable, which work within specific boundaries, and under what conditions they can be used.

A special skill has been created for this. When an agent reaches its first decision point, it goes to the ready system of typical decisions. First it determines which category the question belongs to. If the category is critical, there can be a stop and a wait for me. But there are really few such questions. From my observations, 90% of questions are resolved automatically.

The agent finds several similar questions and the decisions already made on them, and uses this to understand which decision would be closest to mine.

I have already tested this system. I answered somewhere around 40, actually closer to 50 questions, while the agent answered them in shadow mode. Then we compared which answers it had chosen and which I had chosen.

Each question had three answers, and every answer was a real path we could take. We got around 80% agreement. There were subtle differences in the remaining answers. I gave additional explanations, and we refined it.

I see this thing as simply indispensable for any business owner.

Look, you work, you have a team, and every day people can pull you into all kinds of questions. But imagine taking the entire history of questions brought to you and converting it into a knowledge reference: how to act in which situation.

For your team, you can collect thousands of ready decisions that have already been made and package them. And you can imagine different ways of using this.

For example, create a personal assistant. An employee writes to it at any moment and gets an answer. They do not have to wait for when they can call you, when they cannot, whether they should first call your deputy or somebody else.

The assistant has already digitized a large part of your decisions. Say 80% or 90% of typical questions that are not critical, do not concern very large sums, or other things requiring your personal involvement.

And that is it, you get a ready decision in that same second, the business moves faster. Owners and managers are interrupted less, decisions are made faster, and people doing the work feel more independence because they do not have to keep depending on someone and waiting.

We will introduce this into our project too. For now I am testing the pilot through my own experience. We will see, maybe something else will come up, and we will adjust it.