KarnEvil9 Unveils Deterministic AI Agent Runtime Based on Google DeepMind Framework

KarnEvil9 represents a groundbreaking open-source deterministic AI agent runtime, distinguished by its direct implementation of Google DeepMind's Intelligent AI Delegation framework (Tomasev, Franklin & Osindero, 2026). This makes it the first public codebase to translate all five pillars of this influential academic paper into runnable TypeScript. The core innovation lies in its ability to convert natural-language tasks into explicit, structured execution plans, moving away from opaque prompt chains. This deterministic approach ensures that every run is replayable and auditable, a critical feature for accountability in AI systems. A cornerstone of KarnEvil9's architecture is its "accountability by design." It fully integrates DeepMind's delegation framework, incorporating sophisticated mechanisms such as trust scores, escrow bonds, and automatic re-delegation upon task failure. For multi-agent delegation, the runtime employs an impressive suite of nine safety mechanisms: cognitive friction, liability firebreaks, graduated authority, escrow bonds, outcome verification, consensus verification, reputation tracking, delegatee routing, and re-delegation. These mechanisms are designed to govern complex interactions and ensure robust, safe operation. Furthermore, KarnEvil9 features a tamper-evident journal, utilizing a SHA-256 hash-chain event log to detect any post-hoc modifications to the execution history, thereby guaranteeing the integrity of audit trails. Its "domain-ignorant governance" is another significant aspect, meaning the safety layers manage trust and chain depth independently of specific domain knowledge, allowing the same framework to govern diverse workflows from code refactoring to financial transactions without code changes. An illustrative experiment involving a three-node AI swarm playing Zork I demonstrated the framework's capabilities, revealing how initial governance rules could block necessary actions, underscoring the need for trust-aware rather than merely command-aware governance. Requiring Node.js >= 20 and pnpm >= 9.15, KarnEvil9 supports various LLM planners and execution modes, offering flexibility for developers building advanced, auditable, and safe AI agent systems.