Arcee nac:用编排器与线程 episode 撑起长程 Agent 任务的开源 harness
Arcee AI 开源的 nac 将规划与执行拆开:中央编排器只负责任务分解,线程执行后回报结构化 episode,并提供本地仪表盘与多种模型登录方式。
长任务里把所有步骤塞进同一个代理循环,很容易偏离最初目标。该开源项目定位为长程代理运行架,面向实验、训练监管、基础设施与原型等必须对齐原始意图的工作。
按介绍帖与仓库说明,架构采用编排器、线程与回合摘要:中央编排器负责规划与拆解,但不能直接跑命令或改文件;它只启动线程,由线程执行并回报结构化摘要供后续编排。本地可安装网页控制台,在项目目录启动后于浏览器打开仪表盘。模型接入支持设备码登录、主流编程代理的授权登录,以及兼容接口密钥;并提供引导技能与协议服务,方便其他编程代理把长任务委托出去。
安装官方给出脚本将控制台放到用户本地可执行目录。对软件研发读者,其价值在于把持久意图状态与一次性执行上下文分开,减轻长会话上下文腐烂;选型时仍须评估本地守护进程权限、协议暴露面,以及编排器不可直接改文件这一约束是否匹配团队工作流,并确认并行线程的隔离是否满足代码与密钥边界。
nac open-source agent harness for long-running work
Long-running agent work needs more than a bigger prompt.
nac is an open-source agent harness for builders running longer tasks such as experiments, training runs, infrastructure work, and prototyping.
It helps you keep work aligned with the original intent by separating planning from execution: a central orchestrator decomposes the work, while threads execute it and return structured episodes.
Key features:
• Central orchestrator – plans and decomposes work without directly running commands or editing files • Threaded execution – launches threads to handle the work instead of putting every step in one agent loop • Structured episodes – threads return summaries of what they accomplished for the orchestrator to use • Local dashboard – install nac-web, start it from a project, and open the interface locally • Flexible model access – supports Arcee login, ChatGPT Codex OAuth, and OpenAI-compatible provider API keys
It’s open-source (Apache License 2.0 license).
Link in the reply 👇