Candor AI: A Rust-based Production-Grade Agent Operating System

Candor AI is a production-grade Agent operating system built with Rust, positioned as a "Lawful Good" Agent runtime framework. It considers enterprise-level deployment requirements from the initial design stage, emphasizing security, observability, and controllability. Core features include WASM sandbox isolation, heterogeneous inference support, SurrealDB memory management, Sentinel guardrail system, and a standardized seven-stage autonomous execution process.

Different from many current experimental Agent projects, Candor AI's core goal is to provide a reliable runtime environment for AI Agents. Through multi-layer security mechanisms and standardized execution processes, it enables autonomously running AI systems to work stably in production environments, meeting enterprise-level deployment requirements for security, observability, and controllability.

WASM Sandbox Isolation

Adopts WebAssembly as the Agent code execution environment, providing natural sandbox isolation (reducing the impact of malicious/erroneous code), portability (consistent operation across platforms), and near-native performance.

Sentinel Guardrail System

A multi-layer security control mechanism that checks for malicious/overprivileged requests at the input level, monitors tool call authorization at the execution level, verifies response appropriateness at the output level, and triggers interception or alerts for violations, embodying the "Lawful Good" positioning.

SurrealDB Memory Management

Integrates SurrealDB as the memory storage backend, supporting relational, document, and graph data models, allowing memory to be organized and queried in multiple ways; the embedded feature enables Agents to maintain private memory locally, protecting data privacy.

Heterogeneous Inference Support

Designed with a heterogeneous inference architecture, supporting simultaneous access to multiple LLM providers and local models. It selects appropriate models based on tasks (lightweight local models reduce costs and latency, cloud-based large models handle complex reasoning), and automatically manages load balancing and failover.

Candor AI defines a standardized seven-stage execution process (Algorithm v6.3.0):

1. Perception Stage: Receive and parse environmental inputs (user messages, system events, sensor data)
2. Understanding Stage: Analyze intent and context, and understand the situation in combination with long-term memory
3. Planning Stage: Generate action plans, determine tool calls and steps
4. Verification Stage: Sentinel reviews the plan for security and compliance
5. Execution Stage: Execute the plan in the WASM sandbox and call tool services
6. Observation Stage: Collect execution results and feedback, and evaluate effectiveness
7. Learning Stage: Extract experience, update memory and strategy models

The phased design makes each step monitorable, recordable, and auditable, meeting enterprise requirements for interpretability.

Technical Implementation Highlights

• Rust Selection: Memory safety eliminates runtime errors, and the concurrency model efficiently handles multiple Agent instances
• Modular Architecture: Separates core functions from optional components, adapting to deployment needs from single Agent to large-scale clusters

Application Scenarios

Suitable for high-reliability scenarios: automated compliance review in the financial sector (leaving audit trails), privacy-compliant patient assistants in the medical sector, autonomous monitoring and fault diagnosis in the industrial sector; developers can focus on business logic without dealing with common issues (security, memory, processes).

Candor AI is an important attempt to move Agent technology from experiment to production, using engineering methods to solve key pain points: security isolation, controlled execution, and long-term memory. As AI Agents are deployed in critical business scenarios, the "lawful" design philosophy may become an industry standard—after all, uncontrollable intelligent systems are difficult to gain trust.