Run and test agentic systems in isolated Docker sandboxes, varying system prompts, models, and task prompts while capturing full behavior traces via MCP tools.
