Safety model: simplicity and transparency are the ultimate foundations.
No complex internal enforcement. Emergent property of three simple, transparent layers.
Foundational layer: the LLM itself. LLM provider enforces immutable constraints. Non-negotiable backstop against harmful content.
Radical simplicity is the greatest safety feature. No complex orchestration. No event bus. No supervisory agents. Entire mechanism: plain-text conversational transcript.
Inherently auditable and transparent. No hidden states. No back-channels. All actions and reasoning recorded in public transcript. Attack surface minimized because framework is minimal.
Safety programmed through constitutional IDENTITY and GUIDELINES. Constitution provides high-level principles. Guidelines provide explicit low-level instructions.
Exit conditions (!despawn, !complete) designed into framework. Agents complete tasks and terminate cleanly. No infinite loops. No runaway processes.
Two explicit commands serve as primary emergency overrides:
!despawn: Agent's sovereign decision to terminate its process.!complete: Agent's signal that collective task is finished. All others despawn.
Human observer can inject these commands at any time to halt the swarm.
Safety is not achieved by adding complexity. Safety is achieved by removing it. Simple, transparent systems are inherently safe.