AI Does Not Clean Up After Itself

One of the biggest problems in working with AI is the endless growth of everything that can grow: files, research, data, scripts, plans, and documents.

AI still does a poor job of cleaning up after its own work. It is like a builder who came, constructed a new building, but did not wash the tools, sweep the dust, or clean up around it.

Today I restructured this part of the project. After a small morning health check of the whole system, it turned out that the project was adding more than 100 megabytes a day. When a project grows, an increase in size is normal, but daily data growth at that scale already showed that storage was organized inefficiently.

Some parts of the system were saving both aggregated and raw data. Others created a new copy of the entire history every day and added one more package for the current date. After the analysis, there turned out to be quite a lot of such cases.

My new strategic operating system allowed the orchestrator to see the whole project and assess which data was really needed, which form made it easier to use, and what could be rebuilt. Then about six different modules received tasks and reworked their own storage, packaging, and repackaging of data.

The goal was simple: do not accumulate unnecessary raw data, package the necessary information carefully, and store it in a format suitable for the system, backtests, and production.

In practice, this produced one clear data management protocol. It defines how to build new modules, which data to save in them, how to account for it, and which storage methods are unacceptable.

We will see how this works over time and whether the project starts growing quickly again. So far, the growth has been slowed significantly.

The single orchestration layer really showed its strength today. In a few hours, it helped examine a large number of processes, determine what was working efficiently and what was creating waste, and rebuild the data storage.

This also affects how many agent tokens are spent. When an agent sees an enormous number of files, it comes into contact with them one way or another: context breaks apart somewhere, important information gets lost somewhere, and tokens are simply overspent somewhere. The smaller and cleaner the project an agent works with, the more efficiently it works.