WHAT NEW IN X11.03 TUE

X 上的 AI 合集档案 · 22 个合集 · 66,612 条

gpt-6VIDEO

Astra agent 通宵跑 GitHub 仓库的实测栈

作者给 Astra agent 一个仓库跑一夜:检查 73 个 issue、尝试 21 个补丁、16 个测试通过,零改动未经批准合并。

Lummox
@Lummox_eth27 天前收录原推发布于 2026.10.06
Lummox @Lummox_eth原推发布于 2026.10.06
在 X 上打开原推

gpt 6 astra loops are f…cking illegal i gave one Astra agent a GitHub repo overnight and woke up to 73 issues inspected, 21 patches attempted, 16 tests passed and 0 changes merged without approval the stack was basically this : Serena maps the codebase and finds the symbols, files and dependencies Astra actually needs, so it doesn’t burn context reading the entire repo ▸ https://github.com/oraios/serena LangGraph keeps the job moving through plan → edit → test → retry and checkpoints the state when one step fails ▸ https://github.com/langchain-ai/langgraph E2B gives every attempt an isolated sandbox where Astra can run code, install dependencies and break things without touching the real environment ▸ https://github.com/e2b-dev/E2B Langfuse keeps the prompts, tool calls, traces and failures attached to the run, so when Astra makes a bad decision you can actually see where it started ▸ https://github.com/langfuse/langfuse what surprised me is how little Astra needed me once every layer had one job it can understand, code, fail and retry all night but production still gets one final human gate

这条是引用着另一条写的:x.com/promixyz/status/2102424036416348494

x.com/lummox_eth/status/2107485348058030580采集于 2026.10.07 10:10