We've been solving similar problems differently. Instead of a flat progress.txt, we use three layers: semantic context (what are we building and why), structured intention tracking (blockers, dependencies, decision paths), and git for versioning. Separating "understanding" from "tracking" from "versioning" reduces cognitive load on the agent. On the premature victory problem — we lean toward better task decomposition and human checkpoints rather than JSON flags the model "can't" edit. Funny thing: we had a bug where the orchestrating agent couldn't see the sub-agent's completion report. While fixing it, we realized — wait, this is actually useful. The human validates what was actually done before the orchestrator sees it. Accidental feature, now intentional. Feels more robust than programmatic constraints. Anyone else experimenting with multi-session agent architectures? Curious what's working for others.