I get why people are confused

Recently I've been trying to codify some of the things I've been doing with AI to be a little more rigorous. That's when things got weird.

I'm in a polishing phase (mostly) for my latest app Castria. As such I don't want the big loop cycle of directly going back and forth with Claude like I do when we're building the core architecture or distributed systems. I've been using this prompt:

I'd like you to act as an orchestrator for sub-agents. I'll give you tasks and you farm them out to sub-agents, monitor them for completion, when they are done use another sub-agent to verify their work. The key though is that you're returning immediately after dispatching a sub-agent to accept the next task. Rinse and repeat.

When I open a new Claude Code session and use this exact prompt, I get WILDLY different outcomes. Sometimes it works exactly the way I intend, maybe even better than I specified. Other times Claude Opus 4.8 on Ultracode seems very confused about what it's supposed to be doing - ignoring the sub-agent instruction completely.

I'm pretty sure what's happening is that each session is being routed to a different experiment, either slightly different model checkpoints, an attempt to make some part more efficient, etc. where Anthropic is measuring the outputs in order to determine which branches to keep and which ones to prune. But seeing wildly different behavior like that to normal people just cements that this stuff is unreliable and can't be trusted.