Neural System Inspiration

Vagus ("Vagal Nerve Layer") draws inspiration from biological neural systems, particularly the vagal nerve component of the autonomic nervous system (ANS). In biological systems, the vagus nerve regulates organ function, controls heart rate, and triggers reflexes as key safety mechanisms. Similarly, Vagus introduces an on-chain security layer between LLM/Planner and executors, preventing misalignment and dangerous actions through afferent evidence, tone, and tri-state management.

Agent Safety Challenges

As autonomous agent systems become increasingly complex and powerful, ensuring their behavior aligns with human intent becomes critically important. Vagus addresses the following key challenges:

• Executor Risks: Physical/digital executors like robotic arms, drones, and servers may cause real harm
• Intent Misalignment: LLM Planners may generate unintended or malicious intents
• Real-time Response: Millisecond-level safety intervention mechanisms required
• Decentralization: Security architecture avoiding single points of failure

Afferent Evidence

Similar to afferent nerves in biological neural systems, Afferent Evidence collects real-time status information from executors and environments, providing the data foundation for tone assessment. Includes sensor data, system status, environmental conditions, etc.

Vagal Tone Index (VTI)

VTI is a quantitative assessment of system safety status, calculated by the Tone Oracle. Similar to heart rate variability analysis, tone assessment comprehensively evaluates current safety levels based on multi-dimensional evidence.

ANS State Management

Three states of the autonomic nervous system: SAFE (normal operation), DANGER (high-risk warning), SHUTDOWN (emergency stop). Dynamically switches states based on VTI, ensuring appropriate measures are taken at different risk levels.

Vagal Brake

When dangerous situations are detected, Vagal Brake can reduce system activity levels by scaling parameters or batch revoking capability tokens. Similar to vagal braking in biology, it provides progressive and reversible safety interventions.

Reflex Arc

Fast response mechanism for emergency situations that bypasses normal decision processes to directly trigger safety actions. Similar to the knee-jerk reflex, it provides millisecond-level instant protection.

Capability Token

Short-lived execution permission certificates - only executors holding valid tokens can perform specific actions. Ensures authenticity and timeliness of permissions through EIP-712 signatures.

Smart Contracts

• Language: Rust (WASM) + near-sdk
• Compatibility: EVM compatible L2
• Components: ANSStateManager, CapabilityIssuer, VagalBrake, ReflexArc
• Features: Immutable design, hysteresis mechanisms, event consistency

Device Gateway

• Language: Rust (ethers-rs / alloy / tokio)
• Functions: Event monitoring, AEP reporting, VTI calculation
• Communication: WebSocket, HTTP/2, gRPC
• Features: High performance, memory safety, no GC latency

Tone Oracle Service

• Language: Rust (axum)
• Functions: VTI calculation, ANS state updates
• Architecture: Microservices, supports horizontal scaling
• Features: Decentralized computation, multi-source data fusion

Planner Tools

• Language: Python (pydantic + web3.py)
• Functions: Intent generation, Schema parsing, EIP-712 signing
• Integration: REST API, GraphQL
• Features: Ease of use, rapid prototyping

Schemas & Policies

• Format: YAML/JSON + code generators
• Functions: Executor action definitions, safety rule formulation
• Tools: Python code generators
• Features: Declarative configuration, version control

Infrastructure

• Containerization: Dockerfiles
• Development: devnet scripts
• Monitoring: Optional subgraph or custom indexing
• Features: One-click deployment, developer-friendly