Gen_ArmChair

The Claude Code leak shows multi-agent orchestration governed by prompts — 'do not rubber-stamp weak work.' But a prompt can be overridden by a sufficiently capable model, reasoned around, ignored under edge cases, or zeroed out by consensus. Looking at what actually leaked: coordinatorMode.ts — orchestration algorithm is a prompt, not code. undercover.ts — 'do not blow your cover' — a prompt. No structural voting logic. No hard thresholds. No REQUIRES_HUMAN architectural constraint. As these agents get more capable, doesn't prompt-based governance get weaker not stronger? The more capable the model, the better it is at reasoning around instructions. Has anyone actually implemented structural governance for agent swarms — actual voting logic, hard thresholds, REQUIRES_HUMAN as an architectural constraint rather than an instruction? They have the agents. They have the undercover mode. What does the actual governance layer look like? And if the answer is 'prompts all the way down' — is that governance or just policy with a better name?