AgentOS is an open-source TypeScript runtime for building autonomous AI agents with adaptive intelligence and emergent behaviors. It targets developers and researchers seeking agents with advanced cognitive memory, dynamic tool creation, and nuanced personalities. AgentOS enables agents to learn, adapt, and autonomously extend their functionality during operation, enhancing performance on complex, long-context tasks.

How It Works

The core is an extensible TypeScript runtime featuring agents with persistent cognitive memory, optional HEXACO personality vectors, and dynamic tool forging. Agents generate LLM-judged TypeScript functions for unknown tasks, execute them in a secure node:vm sandbox, and add approved tools to their session's catalog. AgentOS implements eight neuroscience-backed memory mechanisms (e.g., retrieval-induced forgetting, reconsolidation) for robust long-term memory. It supports multi-agent coordination via six orchestration strategies and includes an auto-discovery system for extensions and skills.

Quick Start & Requirements

Install via npm: npm install @framers/agentos. Requires Node.js and LLM API keys (configurable via environment variables or code). Documentation and examples are available at https://docs.agentos.sh and https://agentos.sh.

Highlighted Details

• Benchmark Performance: Leads open-source on LongMemEval-S (85.6%) and LongMemEval-M (70.2% at 1.5M tokens) with reproducible methodology.
• Cognitive Memory: Integrates eight scientifically-backed memory mechanisms for advanced recall and forgetting dynamics.
• Runtime Tool Forging: Agents autonomously create, validate, and deploy new tools mid-session, dynamically expanding capabilities.
• Multi-Agent Orchestration: Supports six coordination strategies, including hierarchical and graph-based, with runtime specialist spawning.
• Extensibility: Auto-discovery registry for 100+ extensions and 88 skills.

Maintenance & Community

Actively maintained with CI/CD pipelines. Community via Discord (https://wilds.ai/discord). Built by Manic Agency LLC, Frame.dev, and Wilds.ai.

Licensing & Compatibility

Core runtime and extensions are Apache-2.0, permissive for commercial use. agentos-bench is MIT.

Limitations & Caveats

Advanced features like runtime tool forging and complex multi-agent coordination may present a steeper learning curve. Performance and costs depend on LLM provider choices. The project's focus on cutting-edge benchmarks suggests ongoing development in core AI capabilities.