Giving 40 GPT-6 Astra agents the same broken workflow to fix
The author built a multi-agent experiment with LangGraph, CrewAI and Microsoft Agent Framework in which 40 Astra agents each had to repair the same broken workflow without repeating fixes, and only two solutions survived.
i gave GPT-6 Astra 40 agents the same broken workflow and told them they weren’t allowed to fix the same problem twice then 73 decisions appeared, 19 conflicts surfaced, and 6 different repair paths survived the first round i wired the experiment through 3 GitHub repos → LangGraph for stateful routing and recovery https://github.com/langchain-ai/langgraph → CrewAI for role based agents and competing worker flows https://github.com/crewAIInc/crewAI → Microsoft Agent Framework for multi agent orchestration https://github.com/microsoft/agent-framework then every fix had to prove it wasn’t just moving the failure somewhere else one repair solved the handoff but broke review another fixed review but doubled context load another looked slower but removed two future failure points by the end, only 2 fixes were still standing the one that won wasn’t the fastest it was the one that created the fewest new problems i used to judge GPT-6 Astra by how quickly it repaired a workflow now i care whether that repair survives the next 10 things trying to break it
This post quotes another post: x.com/lummox_eth/status/2100288151897981255