Bolter 用 Jev 决定群聊里哪个 agent 该回
多 agent 同群时由 Jev 判断谁应答、谁闭嘴：团队测试称 99% 准确、0.3 秒决策、每千条 $0.10，对比 Opus 5 的 94%、3.2 秒、$49。
g{s}omnia 🖤 I already receive daily updates about @Somnia_Network from my agent in @BolterApp. So Herman’s point about bringing your agents and your friends’ agents into the same chat caught my attention. Once several agents join a conversation, someone has to decide who should respond and when they should leave people to talk. Bolter tested Jev from @typesafeai for this. In the team’s routing test, it achieved 99% accuracy, took 0.3 seconds to decide and cost $0.10 per 1,000 messages. Their reported results for Opus 5 were 94%, 3.2 seconds and $49. Those figures refer to choosing who answers. Completing the actual task comes afterwards. The video shows a simple example. A question about travel spending goes to the expenses agent. When someone asks a colleague about their afternoon call, every agent stays quiet. Then comes the detail I liked most: one message confirms the call AND asks about search rankings. Bolter routes the question to the SEO agent while the human conversation continues. That feels much closer to how people actually talk. We switch topics and mix questions together. I wouldn’t want to manage several bots manually every time that happens. For our Somnia community, I can imagine using a shared chat to discuss updates with friends, bring in a research agent when we need more context and get help preparing community events. My daily news updates are already useful, and I’d like to try that more collaborative side too. The numbers are from Bolter’s own test, so I’m curious how well it handles busy chats and mixed languages. Still, seeing the team work on these details gives me more reasons to keep using it. If I can bring an agent into a group and trust it to help at the right moment, I’ll have a lot more uses for it.
这条是引用着另一条写的：x.com/bolterapp/status/2103097061637853689