Stop copy-pasting from ChatGPT. I set up a 24/7 daemon running 60 concurrent cloud agents across my repos and my mind is blown.
- 发布
- 09/03 23:15
- 首次发现
- 09/04 18:00
- 类型
- 开发者社区 · RSS
一位开发者指出,许多人错误地使用AI编码工具,仅仅是复制粘贴或将其当作Jira使用。他们分享了自己的高级设置:一个24/7运行的守护程序,拥有60个并发云代理。该系统每45秒检查一次云容量,并立即从JSON积压中提取任务,部署到新的云虚拟机。它能自动将任务路由到特定存储库,例如为前端编写Three.js着色器或处理计费后端的Stripe网络钩子,从而优化了多存储库开发和PR合并流程。
I feel like 99% of people are using AI coding tools completely wrong right now. Most devs I know are either letting Copilot tab complete a function or copy pasting snippets back and forth with Claude in a browser tab. Even people using newer agent tools are basically using them like Jira where they assign one bug wait 15 minutes look at the diff and repeat.
A few hours ago I had a weird realization while hacking on my project. Im building an ecosystem with a CAD CAM engine a 3D web frontend and a billing platform across 6 repos. I have Google Jules Ultra which gives you 60 concurrent cloud VMs and 300 tasks a day and I realized leaving those slots idle when Im not at my keyboard is just wasting compute.
So I put together a simple closed loop system. Locally I use Antigravity right in my terminal as my co pilot and architect to design the specs and math together. In the cloud I use Google Jules which spins up an isolated Linux VM clones the repo installs dependencies writes the files runs tests fixes its own compiler bugs and pushes a real Git branch. To connect them I wrote a PowerShell daemon that runs continuously on my laptop.
Every 45 seconds the script checks my active cloud capacity. If 15 agents finish their tasks and free up slots it instantly pulls the next 15 items from a centralized JSON backlog and fires them off into fresh cloud VMs. It routes tasks to specific repos automatically so one agent writes Three.js shaders in the frontend repo another writes STEP exporters in the core kernel and another handles Stripe webhooks in the billing backend.
The craziest part is the auto hunter fallback I added. If my manual task queue ever empties out the script doesn't idle. It scans the repos for untested files or loose TypeScript any types and spawns agents whose only job is to write Vitest test suites and pay down technical debt until the slots are full again.
I went from spending 12 hours a day manually typing boilerplate chasing syntax quirks and writing unit tests to basically sitting in the cockpit as a VP of Engineering. I set the architectural vision my script keeps dozens of cloud machines coding simultaneously around the clock and I just review the Pull Requests and merge green builds.
When you stop treating AI like a chat assistant and start treating it like an asynchronous headless engineering fleet the leverage is unreal. You're essentially running a 50 person dev team completely solo from a laptop.
Anyone else experimenting with headless agent dispatchers or saturation scripts like this? How are you guys handling multi repo PR merges without losing your mind?