用 Jev 做廉价核查层防止 agent 编造
多 agent 工作流里让便宜模型只判断「结论是否有来源支撑」，用 Jev 跑核查几乎零成本，三轮不过就退回。
Your AI agents are lying to you and you have no f*cking idea Everyone is copying that setup where one dev runs 300 Claude Code agents in parallel. Nobody talks about the part that actually makes it work The secret is the layer that refuses to believe them Here's how it runs: - one model does the work: writes, researches, fills the profile - a second, cheaper model gets one job: is this claim backed by the source, yes or no - no proof = the output goes back. proof = it stays - three passes, and whatever survives lands in the vault The trick is the split. The worker never grades its own homework A single agent will confidently make up a number and move on. 300 of them without a checker is just 300 confident liars Run the checker on something cheap like Jev or DeepSeek Flash and the "verify everything" pass costs almost nothing Same Second Brain + swarm idea everyone is sharing. This is the piece that turns it from a demo into something you can trust
这条是引用着另一条写的：x.com/bober_smart/status/2078784709253841039