How are you handling the orchestration for the Computer Use agent? Is that running on LangGraph or did you roll a custom state machine? I've found managing state consistency in long-running agent loops to be the hardest part to get right reliably.

No custom state machine or agent, it's only a copy of Anthropic's 3 computer use tools: run_bash, edit, computer.

https://github.com/ServiceStack/llms/tree/main/llms/extensio...

It's run in the same process, there's no long agent loops, everything's encapsulated within a single message thread.