About this project

Agent-kernel functions as an operating system kernel for multi-agent systems, managing concurrent AI agents through priority scheduling (CRITICAL to BACKGROUND levels), resource allocation limits for LLM calls and tools, and deadlock detection. It enforces per-agent resource budgets for tokens, time, and tool usage. The message bus supports pub/sub and point-to-point communication. Additional capabilities include rate limiting, concurrency primitives (mutex, semaphore, barrier), health monitoring with auto-recovery, swarm mode for decentralized routing, model routing by complexity, observability metrics, task decomposition with graphing, agent service discovery via A2A registry, and centralized tool management with MCP registry. Written in TypeScript and installed via npm, it includes examples for quick setup and is framework-agnostic.