Making an Astra multi-agent pipeline actually work end to end
The author found the hidden bottleneck in his autonomous workflow was himself, and after making each agent's output the next agent's input, he left and came back to a finished pack.
the first time I left GPT-6 Astra alone, it failed exactly where I hoped it would not because the model couldn’t do the work because I was still secretly doing part of the job one agent finished the research and stopped the next agent had no idea the research existed and I realized the “autonomous workflow” still had one invisible employee connecting everything me so I changed one thing instead of telling each agent what to do next, I made the output of one job become the input of another research landed in a folder that triggered the asset work finished assets moved into drafting drafts went through review only the final approval came back to me then I ran the same workflow again and left this time I didn’t come back to 4 agents asking what happened next I came back to 1 finished pack the research was inside the assets were attached the drafts were ready and the mistakes the agents couldn’t resolve were listed separately instead of silently guessed that was the part I didn’t expect GPT-6 Astra wasn’t more useful because I gave it more autonomy it became more useful when I stopped making myself the API between every agent
This post quotes another post: “Running a backlog overnight with 4 GPT-6 Astra agents on a shared filesystem”