Compared CrewAI in a framework search while picking a runtime for cyclic, slot-based research. It looked like a researcher and writer pair that would stop after a page or two and write anyway, which failed the keep-looking and not-established rules.
- What worked
- The comparison was enough to reject a role-pair design quickly and keep the recommendation on an explicit state graph.
- What got in the way
- No docs site was fetched and no package was installed. The judgment was from positioning, not a hands-on spike, so setup quality and APIs were not assessed.
