Grok Bot 团队工作流里补上 Jev 决策层

silentguy @silentguyy66原推发布于 2026.09.29

在 X 上打开原推

SpaceXAI just dropped a free workshop 55 minutes on running a team of Grok Bots that merges its own PRs while you sleep: 01:36 - her agents automerged 20 PRs overnight and she stopped reading the code 31:18 - grok bot launches: every agent has its own identity and you orchestrate them like people 37:01 - dune architecture: useEffect banned, CI turns the build red and yells at the agent 50:29 - CI goes red so the agent can't just write crappy code 53:10 - designers and PMs now ship grok bot features without writing code 600+ PRs of refactoring is what it took to get a codebase her agents could trust Guardrails, banned imports, green build or it doesn't land But every rule she encoded is a hard-coded one A team of agents still has to answer four soft questions before each move Her stack sends all four back to the frontier model, which answers in a paragraph of JSON and hopes the schema survives Jev is the layer that was missing It never writes, it decides Typed answers plus a probability your code can route on Grok Bot creates the work → Jev decides what happens next → code enforces the decision Twelve agents without it are twelve tabs pretending to be a company Save this before you spin up another agent

这条是引用着另一条写的：「Jev 把路由打分验证做成决策层」