An Astra agent ran a GitHub repo overnight: 73 issues, 21 patches, 16 tests passed
The author gave one Astra agent a GitHub repo overnight and it inspected 73 issues, attempted 21 patches, passed 16 tests, and merged no changes without approval, using Serena, LangGraph, E2B, and Langfuse.
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
This post quotes another post: x.com/promixyz/status/2102424036416348494