Running My Own Agent Runtime
- Part 1A third of the window was already full before any work started
My AI assistant kept breaking its own rules. The cause was not the model, it was the documents I wrote. 63KB loads on every session, and the rules were buried inside it.
- Part 2My markdown folder was already a multi-agent system
Running five AI sessions in parallel, records kept vanishing and state kept drifting. Chasing the cause, I realized my canonical markdown files were the only communication channel between agents.
- Part 3My AI erased my records twice
Overwrite a file from a stale snapshot read early in the session and everything another session appended in between disappears. It happened twice, and the second time it became a rule.
- Part 4You don't invent the answer. You converge on it.
When direction is unclear I stopped thinking harder and built a loop that keeps researching instead. The first time I ran it for real, the loop disproved a premise of my own goal.