I now have a second job of verifying the janky software that agents built for me. And I'm paying two subscriptions for the privilege.
The context-switching and constantly having to reverse-engineer why agents did what they did was getting too exhausting, so I built a thing about it.
Now I plan my work in small vertical slices up front and store all the necessary context in a plain-text knowledge base repo, with links so agents can discover what they need.
Then I can cold-start an agent session with a prompt generated by my tool, so the agent has a clear brief and can sue the tool to write its progress back to the kb.