building a solo game with 9 AI-assisted workstreams. how would you actually orchestrate them?

I'm working on a 2D co-op arcade game, mostly as a one-person developer using AI for a lot of the implementation. The project has grown into nine different workstreams: assets, animation, audio, camera/rendering, local co-op, QA, enemies/combat, levels, and save/replay. And I'm starting to wonder how people actually coordinate something like this without becoming a full-time manager of AI conversations. To clarify, these aren't nine autonomous agents already running in the background. They're separate areas of work, with their own experiments, Git branches, specifications, and development conversations. Some components work independently, but the complete production pipeline isn't integrated yet. The direction I'm working toward is to give each specialist a complete objective, let it implement and test within agreed boundaries, and have it return code, assets, test results, captures, and a summary of what changed. Important decisions should live in Git and project files, not in chat memory. The human still reviews creative quality and approves major changes. But here's where I get stuck. Suppose the animation workstream changes the timing or position of a weapon muzzle. That could affect the combat code, rendering, audio synchronization, and tests. Who detects that dependency? Who decides which specialist needs to do what? And who makes sure the final combined result actually works? I'm considering a small durable job system with dependencies, structured outputs, checkpoints, and independent validation, rather than a bunch of agents continuously messaging each other. For people who've built something similar: How much of the orchestration should be ordinary deterministic software, and how much should an LLM decide? How do you manage dependencies, retries, incomplete work, human approvals, and switching between AI providers without losing context? And perhaps most importantly, what's the smallest architecture you've actually used successfully for this kind of setup? I'm not looking for a magical framework that promises a fully autonomous company. I'd really appreciate hearing what has worked in real projects, what failed, and what you'd do differently.