I went into a bunch of OpenClaw discussions expecting the usual advice about subagents: better prompts, cleaner folders, maybe some heroic config. What I found was more interesting. The OpenClaw setups that actually seem to hold up are not just "one agent with more prompts." They are separate services with separate trust zones. The pattern that keeps showing up looks like this: a librarian agent an executor agent a company-facing agent Usually connected over A2A. That sounds like a small implementation detail. It is not. A separate prompt inside one workspace is still one workspace: one context blob one tool surface one security boundary one place for bloat to accumulate A separate OpenClaw instance is different. Now you have real boundaries: different runtimes different API keys different networks different memory policies explicit handoffs That is where multi-agent starts being architecture instead of roleplay.…