A developer built a 10-agent production team using OpenClaw, utilizing Soul.md for personas and Convex as a shared real-time brain. The system relies on file-based Memory (WORKING.md) for task continuity and 15min heartbeats to keep Cost manageable. Specialized Roles (e.g., skeptical analysts) proved more effective than generalists.
Does file-based state beat vector DBs for active agent coordination? How would you handle persistent memory across long-running autonomous sessions?