Week 0x2C

Still more work on focus and sandbox execution this week. I vibed a bit too much and lost track of what was being done and we burned way too much time and tokens essentially reproducing the sandbox working model in the ACP adapter. All of that just got thrown away, but at least the overall plan matured a bit.

Part of that was how to hold an agent loop accountable for doing work. My hypothesis is that a simple deterministic check for any filesystem changes or simply asking the model for a reference to the output will be more than enough to ensure work is performed.

Along those lines, we’ve introduced a journal for autonomous jobs, where the model can leave notes for itself, log decisions, etc.