A weekend building agent systems with GPT-6 Astra: the real problem was wiring outputs together
The author built a multi-agent workflow with Astra and found the real issue was not connecting the outputs, which led to paying twice for the same research.
i spent an entire weekend building with GPT-6 Astra before realizing i was the reason the system kept forgetting everything at first, it worked then it got weird every morning, more work was finished the agents were researching, connecting ideas and organizing files while i was somewhere else. i thought i had finally built the system i wanted until i caught myself doing the same thing again i remembered one useful detail and spent 20 minutes digging through old folders trying to find it then i found something worse i had paid for basically the same research twice same question almost the same conclusions two separate runs that had no idea the other existed that was when i stopped building more agents for the next few days, i connected what they had already produced. research stayed attached to companies, decisions stayed connected to sources, and failed experiments stopped disappearing slowly, the workspace started turning into what you see in this video some branches kept growing others stopped then i opened a completely new task and realized half the work was already there not because Astra magically knew the answer because i had already paid to learn it once i spent months making AI better at doing my work turns out the bigger upgrade was making sure neither of us had to do the same work twice
This post quotes another post: x.com/lummox_eth/status/2100288151897981255