Skip to content

Open-source TypeScript AI agent framework

AgentOS is an open-source TypeScript AI agent framework for building production-ready autonomous agents with multi-agent orchestration, cognitive memory, multimodal RAG, AI guardrails, voice pipelines, and 13 LLM providers. A type-safe alternative to LangGraph, CrewAI, Mastra, and AutoGen.

기억하고, 적응하고, 새로운 도구를 만들어내는 AI 에이전트를 위한 오픈 소스 TypeScript 런타임. 전체 런타임의 각 세션은 GMI(Generalized Mind Instance)로 실행되어 고유한 페르소나, 무드, 메모리를 유지하고 시간이 지남에 따라 자신의 특성을 조정할 수 있습니다. 에이전트는 작업 도중 직접 함수를 작성하고 심판이 이를 승인하며, 고정 로스터가 작업을 감당하지 못하면 매니저가 전문가를 생성합니다. 에빙하우스 망각 곡선과 7개 슬롯의 작업 기억을 갖춘 인지 메모리. LongMemEval-S 85.6%. Apache 2.0.

See code examples
—GitHub 스타
—포크
PII 마스킹 가드레일 팩자체 인프라에 셀프 호스팅

Runtime agent spawning

Spawn an agent at runtime

A manager coordinates a researcher and a writer. When the task needs a role neither covers, the manager calls spawn_specialist, an LLM judge approves the new agent's spec, and the specialist joins the roster for the manager's next turn.

import { agency } from '@framers/agentos';

const research = agency({
  provider: 'openai',
  model: 'gpt-4o',
  strategy: 'hierarchical',
  instructions:
    'Coordinate the team to produce a thorough research report. ' +
    'If you need a capability neither the researcher nor the writer covers, ' +
    'call spawn_specialist to mint a new specialist for just this run.',
  agents: {
    researcher: { instructions: 'Find authoritative sources and pull verbatim quotes.' },
    writer: { instructions: 'Produce well-cited prose from the research brief.' },
  },
  emergent: {
    enabled: true,     // unlocks the spawn_specialist tool
    tier: 'session',   // spawned agents are discarded when generate() returns
    judge: true,       // EmergentAgentJudge reviews each spec before activation
    planner: { maxSpecialists: 3, requireJustification: true, maxJudgeCalls: 6 },
  },
  on: {
    emergentForge: (event) =>
      console.log(`[FORGE] spawned "${event.agentName}" approved=${event.approved}`),
  },
});

const result = await research.generate(
  'Survey the post-2023 RAG literature, including evaluation methodology, ' +
    'known limitations, and reproducibility challenges.',
);
console.log(result.text);
Condensed from examples/emergent-hierarchical-spawning.mjs. Run the full file with node in a built clone.One captured run →View source on GitHub →

AI Agent Swarm Simulation Engine

Define a scenario as JSON. Run it with AI commanders that have different personalities. Watch their decisions compound into divergent civilizations from identical starting conditions. Built on AgentOS.

See AgentOS in action

Watch a live AgentOS agent handle a real workload end-to-end: tool use, memory recall, multi-step planning. Every clip is captured from a running runtime, no edits.

MIT LicensedOpen Source

AgentOS Workbench

A React + Vite dashboard you run on your own machine to inspect and debug AgentOS sessions. Watch streamed output on a session timeline, compose and replay turns, and run evaluations against your agents.

  • Session inspector for streamed text, tool calls and errors
  • Real-time testing and debugging tools
  • Views for agency strategies, plan checkpoints and RAG retrieval
View on GitHub

Clone the repo and follow the README's quick start; it opens at localhost:5175.

적응형 AI 에이전트를 구축할 준비가 되셨나요?

AgentOS로 차세대 지능형 시스템을 구축하는 개발자에 합류하세요.