AgentOS

AgentOS is a TypeScript framework for building production-ready AI agents that can remember context, coordinate with other agents, and generate new tools on the fly during a task. It's built by Frame.dev and licensed under Apache 2.0, with agents and extensions under MIT.

The memory system is grounded in neuroscience principles and scores 85.6% on LongMemEval-S, placing it within 0.4 points of a leading closed-source SaaS benchmark. Agents don't just recall facts. They adapt their behavior based on what they've learned across sessions.

Runtime tool generation is a standout feature. When an agent encounters a task its existing tools can't handle, it writes a new function mid-task. A judge component reviews and approves the generated code before it runs. Manager agents can also spin up specialist sub-agents when the current roster doesn't cover the work.

The framework supports 21 LLM providers, multimodal RAG, voice pipelines, and GDPR-ready guardrails. It's self-hostable on your own infrastructure. A visual desktop workbench (coming soon) will allow drag-and-drop workflow building and real-time debugging without touching code.

Among AI agent platforms, it's positioned as a type-safe alternative to LangGraph-style frameworks like Agno, CrewAI, and Letta, targeting TypeScript developers who want strong typing and production reliability without sacrificing flexibility.