PLAT-214: Agent Runtime Safety Controls¶
Statement¶
When the platform enables LLM-backed or agentic execution inside ranges, it shall provide runtime safety controls including an operator kill switch, configurable autonomy levels, command or tool allow/deny policy, rate limiting, and approval gates for actions classified as dangerous by the deployment policy.
Rationale¶
Shifter already records command-execution safety as an architecture NFR and has per-range LLM access requirements. LilRAE (formerly APTL) adds the product-level controls needed when agents can act through tools during experiments or live ranges.
Traceability¶
- DOCUMENTS → SPEC
aptl:SAF-001(LilRAE specification, former APTL identifier SAF-001: Kill Switch for All Agent and MCP Operations) - DOCUMENTS → SPEC
aptl:SAF-003(LilRAE specification, former APTL identifier SAF-003: Tiered Autonomy Levels) - DOCUMENTS → SPEC
aptl:SAF-004(LilRAE specification, former APTL identifier SAF-004: MCP Command Filtering (Allowlist/Denylist, Rate Limiting)) - DOCUMENTS → SPEC
aptl:SAF-006(LilRAE specification, former APTL identifier SAF-006: Dangerous Action Approval Workflow)