When Autonomy Starts to Break Down

Autonomy is a flexible concept. It works up to a certain point, and then it starts to break down. I have been through this cycle more than once.

Here are the typical problems I encounter. Autonomous cycles become less effective and degrade each time. They turn into bureaucracy, where one agent carries notes to another, that one carries them to a third, and they keep passing everything around endlessly, shedding responsibility for making key decisions.

This can be anything from managing the resources of my own server and ChatGPT subscription limits to actual spending, APIs and everything around them. Getting them to simply make decisions is a real pain. Even my automated decision-making does not always work. I have to supplement it with rules in AGENTS.md, create ready-made cases where an agent takes a task and goes ahead with it, and give agents full authority. Without some authority they immediately try to step away from responsibility, and the goal gets blocked.

With all the panic around AI being able to do everything, my understanding is that the major labs are tightening the rules and requiring human intervention in key areas: account access, actual spending, limits, architectural changes and so on. And this is where autonomy starts to fall apart.

I have put some things in code and I am still working on others, but it has come down to my daily cycle performing worse and worse over the last week, requiring more and more of my time.

For the last two days I have been doing practically nothing else. My processes, my sockets, continuously listen to a large number of exchanges and record trades, liquidations, open interest and options. All of this is continuously recorded and calculated, but the processes started malfunctioning, and I had to rework the whole cycle.

It used to take, probably, about two hours on average. The last few times the full cycle took around five hours and took me away from everything else. Over the last three days I decided to rewrite the cycle almost from scratch, using the work already done. The goal is to get it down to 30 minutes without losing quality, keeping the gap between the data received and the resulting forecast as small as possible.

There are still self-improvement cycles ahead, and they need repeated adjustments, because forecast accuracy is very difficult here. It is also very important to define boundaries for agents: where one can improve internal calculations, for example, while leaving the external results, design, buttons and everything a person interacts with on the other side of the screen alone.

For me, Crash Brief is an amazing experience that I can actually carry over into other projects, especially where there are people who can support autonomous systems like these.