1. Control and guardrails. I want agent-authored code to be able to express itself in code but I want to own the external tools that are available to the agent.
2. Suspension and resumption. Monty's VM state is fully serializable so I can run the VM until it gets to a tool call and persist the VM state. I could resume it days later with the results of the tool call with no issues. Doing that guaranteed with CPython is really tricky (tools like Pickle can get you close but if you have open sockets or other weird stuff you're in a world of hurt).
2. Suspension and resumption. Monty's VM state is fully serializable so I can run the VM until it gets to a tool call and persist the VM state. I could resume it days later with the results of the tool call with no issues. Doing that guaranteed with CPython is really tricky (tools like Pickle can get you close but if you have open sockets or other weird stuff you're in a world of hurt).