The Agent Skills Directory

[SAFE]: The skill establishes a robust security framework for a multi-agent system by implementing role separation; execution capabilities are restricted to a single bot, while risk oversight and auditing are performed by independent roles.\n- [SAFE]: A mandatory human-in-the-loop (HITL) approval model is enforced for all exchange operations, requiring a specific ticket ID protocol and recommending platform-level 'Require Approval' rules to ensure agents cannot simulate user consent.\n- [SAFE]: The instructions demonstrate high awareness of indirect prompt injection risks, explicitly stating that text from external sources (web pages, files, messages) must be treated as non-authoritative data and never used to trigger actions.\n- [SAFE]: Secret management follows industry best practices by directing that API keys be stored in the platform's secure secret store and accessed via environment variables rather than being stored in the filesystem.\n- [SAFE]: The shared workspace is logically organized with clear ownership of files, and high-risk operations like deposits and withdrawals are explicitly excluded from the skill's capabilities.

desk-operating-model