Problems I'm seeing: - spending 80% of time troubleshooting the agent framework, vs doing constructive projects - agents writing without reading first - concurrency issues, agents reading from stale memory - agents not consistently backing up before chantes - no built-in safety controls for rolling back to last good config - intallers or plugins stomping all over the file - schema constantly changing and agents not checking docs and then hallucinating syntax - basic problems with json (no comments, no types, very fragile to syntax errors) - even small errors will block gateway from restarting and block the agents from being able to run to autonomously troubleshoot or fix their own mistakes. - openclaw doctor seems to cause as many problems as it fixes
I find myself setting up a lot of scripts, scaffolding, and external AI tools, just to babysit an agentic framework. Which sort of defeats the purpose of using a "ready-made" framework in the first place? Or is there an easier way?