NanoClaw V2: Human-in-the-loop approvals and Agent to Agent collaboration

NanoClaw V2 is a comprehensive rewrite of its predecessor, designed to enhance agent capability while maintaining robust safety measures. It introduces human-in-the-loop approvals via messaging apps and persistent agent-to-agent communication, effectively balancing the tradeoff between capability and security. Unlike traditional frameworks that struggled to safely handle sensitive data and actions, NanoClaw employs a sandboxing approach at the agent level rather than individual tool calls, allowing agents to operate freely within isolated environments. This is achieved through the use of Docker sandboxes and OneCLI vaults, which manage credentials without exposing them to the agents, and approval mechanisms that ensure accountability by attaching human identity to sensitive actions. The architecture divides agents by sensitivity, with research agents having internet access and action agents handling sensitive data without internet, communicating through approval-gated exchanges. The system focuses on ergonomics, both for agents and humans, by reducing friction and integrating approval processes into existing workflows like Slack or Teams. This flexibility allows organizations to tailor security and capability settings to their specific needs, challenging the traditional tradeoff between safety and value that OpenClaw's approach represented, and presenting a new paradigm where agents can be both highly secure and highly functional.