What this case covers
The author has a development background and shifted from individual OPC to leading a team of more than ten AI developers. Within half a year, they experienced OpenClaw shutdown, muddled Token accounts, account bans, and permission incidents. The article covers tool selection, account pool solutions, security experiments, and onboarding experience for newcomers.
Questions it answers
- Why did the OpenClaw multi-Agent solution shut down after one month?
- How did the New-API relay station solve the Token allocation problem?
- What pitfalls were encountered in selecting Cursor, Claude, and Codex?
- What risks did the relay station security experiment expose?
- How should veterans run MVPs and newcomers take over maintenance?
Our observations
- All-purpose Agents cannot beat specialized development tools in real engineering delivery.
- Only after Token management is observable can optimization and reimbursement be discussed.
- AI suggestions are only locally optimal; preventing global collapse depends on permission isolation as a fallback.
From the author
First make behavior observable, then talk about optimization - this principle holds for managing Tokens, and it holds for managing other things too.