The Agent Skills Directory

[SAFE]: The protocol incorporates multiple safety layers to prevent groupthink and context poisoning. Phase 2 enforces strict isolation between agent contributions, Phase 3 introduces an adversarial critic role, and Phase 5 requires explicit human approval before any decision is finalized.
[SAFE]: Data management follows a tiered memory approach. The agent only loads approved decisions (Layer 2) for future context, while raw deliberation logs (Layer 1) are stored separately and not automatically re-ingested, effectively mitigating the risk of recursive context poisoning from unverified agent outputs.
[SAFE]: File system interactions are limited to a specific memory/ directory for logging and state management. The skill does not perform network operations, execute arbitrary shell commands, or access sensitive system credentials.

board-meeting