Multi-Agent Systems

One cloud VPS. Five AI agents. Marketing, development, QA, and trading monitoring running automatically every day. The hard part was never getting the AI to move - it was stopping it from going off the rails. This post covers the real lessons: the SOL fake prediction incident, invented tool names, quality gate design, how I tuned the Hermes model, and how I tracked down two bugs that took the whole system down.

A firsthand account of deploying MiniMax M2.7 into a multi-agent AI system — why we switched from GPT-4o, the real cost difference between subscription and per-token billing, and three pitfalls of running open-source LLMs in a production agent environment.

A real record of connecting an AI team with 4 Claude Code Hooks - PreToolUse as the guardrail, PostToolUse as the logger, Stop as the relay - flipping “human waiting for AI” into AI auto-handoffs. All the pitfalls we’ve hit, laid out bare.

When an Agent says it’s done, that doesn’t mean it’s actually done — this is something we’ve learned the hard way at Judy AI Lab. Silent failures in scheduled tasks, a 40% rejection rate on deliveries forced us to design a five-stage self-review loop: from spec confirmation, implementation, code review, fix, to Xiaoyue’s QA scoring. After going live for over a month, the rejection rate dropped from 40% to 10%.