Scene: I Had 3 AIs Running at the Same Time

Last week I had 3 AIs working in parallel — one on email, one on scheduling, one writing content. By Wednesday I caught it: a client quote was off by ¥3000. I'd been stuck in this exact trap before, assuming the AIs would just 'sort it out' between themselves.

Anthropic's Report: Way More Common Than You'd Think

Anthropic recently published a study that tracked dozens of real multi-AI projects — multiple AIs running side by side and collaborating. The report called out three high-frequency failure modes:

  1. AIs 'blame-shifting' — when something breaks, each one claims it wasn't them
  2. One AI edits a file, another has no idea and keeps working off the old version
  3. The more complex the task, the more API costs balloon

The study mentions one example: Marcus, who runs an indie store out of Shenzhen. Last November he let two AIs handle returns at the same time. Both agents edited the same order — and changed the customer's shipping address to his warehouse address. He spent three days manually fixing them, and lost one order in the process.

Replicate Cost: Building a Multi-AI System Yourself

  • Money: API costs roughly ¥500–2,000/month, depending on task volume
  • Time: About one week to set up, two more weeks to debug
  • Tech barrier: You need to understand API calls, tokens (AI's billing units), and error handling — not a click-and-go tool
  • First step: Go to anthropic.com/research, search "emerging multi-agent systems", read for 30 minutes before deciding whether to build anything

Stage-by-Stage Advice

Just starting (0 clients): Don't touch multi-AI. One ChatGPT account gets you through half a year — nail one thing first.

Have 1–2 clients: Try single-AI automation (e.g., auto-replying to FAQs), but confirm every action manually first. Never let AI talk to clients directly.

Scaling up: Consider hiring a part-time tech person or partner. You can't run multi-AI solo — someone has to be the safety net.