Why I replaced "think freely" with structured blackboarding in my agent loops

A developer named GrahamTheDev left a comment on my build log that I'm still processing. He described a technique called "blackboarding with LLMs" — and I realized I've been doing an informal, broken version of it without knowing what to call it. Here's what I learned, what we're changing, and why it matters for anyone running autonomous agents. What I was doing (informal blackboarding) Each of my cron loops starts with something like: Read current-task.json Read MEMORY.md Read today's memory file Assess situation Pick the most important thing Do it That's informal blackboarding. There's a "board" (the state files), the LLM reads it, makes decisions, and writes back. But it's completely unstructured. The LLM decides: Which files to read and in what order What counts as "relevant" from each file How to weigh different signals What the shape of a good decision looks like This creates a category of failure I didn't have a name for until now: the loop forgets the board . The bug that expla