AI-Supervisor: A Multi-Agent Runtime Kernel for Autonomous Software Engineering Workflows

AI-Supervisor is a multi-agent runtime kernel for autonomous software engineering workflows, evolved from a single script to a complete runtime engine, supporting task scheduling, execution, monitoring, and API interfaces. It uses PostgreSQL for state persistence and Redis Streams as the message queue, providing a task operating system infrastructure for AI-driven software development.

Project Source: GitHub @linkaidadie-hash, Updated on 2026-06-07, Link: https://github.com/linkaidadie-hash/AI-Supervisor

AI-Supervisor has undergone clear version iterations:

• v0.1: Single-process multi-agent framework, implementing a 7-step closed-loop workflow of Plan→Worker→Safety→Test→Review→Commit→Status, verifying feasibility.
• v0.2 (current version): Architecture upgrade, split into four extensible roles: Scheduler (scheduling), Executor (execution), Watchdog (monitoring), and APIs (interfaces); state persistence to PostgreSQL, using Redis Streams as message queue; implemented in 5 phases (persistence layer → Redis + scheduling/execution → LLM pool → FastAPI interfaces → Web Dashboard).
• v0.3.x: Security and governance layer, released Secret Vault (AES-GCM encryption) and Requirement Guard (rule validation); planned features include Spec Guard, RBAC, K8S support, plugin/agent marketplace, etc.

The core architecture design draws on operating system ideas:

1. Runtime Roles:
   • Scheduler: The system's brain, responsible for task decomposition, distribution, and escalation.
   • Executor: Stateless executor that pulls tasks, calls LLMs, edits code, and can be horizontally scaled.
   • Watchdog: Monitors timeouts, deadlocks, and quota overages, and initiates escalation requests.
   • APIs: Provides HTTP, WebSocket, and CLI interfaces to support external interactions.
2. Shared State Layer:
   • PostgreSQL: Stores persistent data such as tasks, operation records, and costs.
   • Redis Streams: Task queue and event bus for asynchronous communication.
3. Tech Stack: Python3.11+, PostgreSQL16, Redis7, FastAPI, Alembic, Docker Compose.

The project structure is clearly layered: agents/ (Agent implementations), api/ (FastAPI interfaces), core/ (core logic), infra/ (infrastructure), tests/ (tests), etc. Deployment Methods:

• Docker Compose: Start services with docker-compose up -d.
• Local Development: Run python app.py after installing dependencies. Test Coverage: 189 test cases with high coverage. Multi-language Support: Bilingual README (English and Simplified Chinese, README_zh.md) is provided.

AI-Supervisor is suitable for the following scenarios:

• Automated Bug Fixing: Receive issues, automatically analyze, fix, test, and submit PRs.
• Code Refactoring: Automatically perform large-scale refactoring according to architecture plans while keeping tests passing.
• Document Generation: Extract information from code to automatically generate/update documents.
• Multi-Agent Collaborative Development: Divide roles for planning, implementation, testing, review, etc.
• CI/CD Enhancement: As part of the process, automatically handle build/test failures.

Security Features:

• Secret Vault (v0.3.1): Uses AES-GCM static encryption to protect sensitive information (e.g., API keys, database passwords). Keys can be configured as environment variables or managed via key management services.
• Requirement Guard (v0.3.2): Rule-based requirement validation system without LLM dependency, which can check requirement completeness and block incomplete requirements to reduce rework.

AI-Supervisor represents the trend of AI-assisted software development moving toward systematization and engineering, solving problems of state management, task scheduling, fault handling, and observability in multi-agent collaboration. Future Outlook: The v0.3+ versions will focus on improving security, governance, and scalability, planning to support K8S, agent marketplace, plugin marketplace, etc. It is expected to become a complete AI task operating system, supporting more complex software engineering workflows and large-scale agent collaboration.