Week 0x30

Now I’m getting angry at the AI.

I’m in my third week trying to resolve issues with reloading a session, and Builder Bear continually declares success when the issue has not been solved. But I think we finally found it, after I micromanaged diagnostic logging.

Meanwhile, we also made some progress on the shift from interactive to task-focused loop control, but this, too, required micromanagement. More than anything, it meant removing a bunch of complexity that had accreted over time, as the agents love to add more logic rather than step back and consider if the state machine is well-modelled.1

I also made some initial progress on Cabinet, the knowledgebase, and Github authorization management. Neither is fully useful, yet, but I needed to move forward on something other than “make the loop work reliably”.

Funny prompt of the week:

Let’s update the docket task tool instructions to encourage models to do the same sort of reflection about tree structure that you just performed.