The Agent Skills Directory

[PROMPT_INJECTION]: The skill architecture described in SKILL.md is designed to ingest and process external 'workload' and 'workload_profile' data to direct the behavior of 100+ agents. This creates a surface for indirect prompt injection if the processed data contains malicious instructions.
Ingestion points: Untrusted data enters the agent context through the orchestrate_massive_workload(workload_profile) and coordinate_tenant_workload(tenant_id, workload) methods.
Boundary markers: The design incorporates IsolationManager and TenantManager to delimit execution contexts and verify isolation, providing some protection against cross-tenant interference.
Capability inventory: The coordinator possesses extensive capabilities including global resource allocation, hierarchical agent control, and task execution management.
Sanitization: The implementation logic includes explicit calls to verify_isolation and permission checks (get_info) to validate the context before and after coordination activities.

multi-agent-coordinator