Skip to content- Sources: primary
- Summary: Mislav Ivanda compares three isolation models for agent code: command allow lists with approval prompts, embedded runtimes such as V8 or QuickJS, and a separate VM. He argues that allowing a command like python -c grants arbitrary capability, and that embedded runtimes isolate generated glue code while its tool calls still run on the host. He cites CrewAI removing its local code interpreter in April 2026 and AG2 retiring local executors, and says the execution boundary is harder to swap later than the framework.
- Why it matters: Pick the agent execution boundary before the framework, because tools written for host execution are costly to move into a sandbox later.