I use agents for software work, writing, review, and research, but I do not treat a chat session as a project memory.

If an agent changes a repository, the relevant material has to move back into the repository: files, notes, assumptions, validation, and review. Otherwise the work stays trapped in a transcript that is difficult to reopen and easy to misread.

This is the part of agentic work I care about most. Not autonomy as a spectacle, not demos pretending to be systems, not chains of tools running without context. I care about the layer where work becomes inspectable: source files, state, reports, pull requests, tests, and human decisions.

The chat can help think. It can compare options, draft alternatives, find inconsistencies, and expose weak assumptions. That is useful. But the project is still the repository, the documents, the code, the checks, and the decisions that remain after the chat is closed.

The more powerful the assistance becomes, the more important this layer gets. Generated work without a trace is not faster engineering. It is future confusion written at high speed.