AgentOS
AgentOS is the highest layer in the four-layer capability architecture of agent engineering. It is responsible for multi-agent collaboration and system-level orchestration, uniformly scheduling the execution, communication and resource allocation of multiple agents.
Timeline (last 90 days)
If shared memory has no permission layering, erroneous experience written by one Agent may contaminate the entire system, so before writing to long-term memory it must go through a verification chain and be labeled with its scope of applicability.
When output deviates from requirements, check the Prompt layer; when a task breaks off, check the Context layer; when data transmission goes wrong or the process cannot be recovered, check the Harness layer; when multiple Agents wait for each other, check the collaboration scheduling layer.
The verification chain for experience consolidation is: after the task is completed, first record the execution conditions, process, and results, then extract candidate experience with scope of applicability and evidence, and finally verify it in a new task; only experience that withstands repeated verification is written into long-term memory.
Ewan is building an application called AgentOS on Agent Harness for assigning and tracking tasks for both himself and his agents.
AgentOS has taken over approximately 95% of Danny's programming and operations tasks, including developing an entire web game.
Thanks to built-in end-to-end testing, PRs produced by AgentOS work directly 99% of the time.
AgentOS achieves folder-level and operation-type-level permission control (e.g., write-only, no deletion) by adding an MCP layer on top of Cloudflare R2, with the MCP server validating and intercepting unauthorized operations.
AgentOS 的自动化分为 Tasks（看板式流程）和 Goals（开放式甘特循环）两大部分
计划审查环节由审查协调员派生出四个审查智能体（可行性、范围、守护、一致性），各自出具报告后由协调员汇总
独立开发者 Danny Postma 基于 Anthropic 的 Claude Agent SDK 搭建了一套名为 AgentOS 的智能体操作系统
8 more timeline events
All Facts (18)
When output deviates from requirements, check the Prompt layer; when a task breaks off, check the Context layer; when data transmission goes wrong or the process cannot be recovered, check the Harness layer; when multiple Agents wait for each other, check the collaboration scheduling layer.
50%UnverifiedIf shared memory has no permission layering, erroneous experience written by one Agent may contaminate the entire system, so before writing to long-term memory it must go through a verification chain and be labeled with its scope of applicability.
50%UnverifiedThe verification chain for experience consolidation is: after the task is completed, first record the execution conditions, process, and results, then extract candidate experience with scope of applicability and evidence, and finally verify it in a new task; only experience that withstands repeated verification is written into long-term memory.
50%UnverifiedAgentOS has taken over approximately 95% of Danny's programming and operations tasks, including developing an entire web game.
50%UnverifiedAgentOS achieves folder-level and operation-type-level permission control (e.g., write-only, no deletion) by adding an MCP layer on top of Cloudflare R2, with the MCP server validating and intercepting unauthorized operations.
50%UnverifiedThanks to built-in end-to-end testing, PRs produced by AgentOS work directly 99% of the time.
50%UnverifiedAgentOS 包含多个分工明确的专精智能体（规划智能体、资深开发、审查协调员等），每个都有专属prompt、技能集、MCP连接和知识库
50%UnverifiedAgentOS 的自动化分为 Tasks（看板式流程）和 Goals（开放式甘特循环）两大部分
50%Unverified计划审查环节由审查协调员派生出四个审查智能体（可行性、范围、守护、一致性），各自出具报告后由协调员汇总
50%Unverified独立开发者 Danny Postma 基于 Anthropic 的 Claude Agent SDK 搭建了一套名为 AgentOS 的智能体操作系统
50%UnverifiedAgentOS 通过 Inbox（本质是一个MCP）实现人机协作，支持开放式问答和带选项的多选题，作为渐进式Web应用（PWA）部署
50%Unverified得益于内置的端到端测试，AgentOS 产出的 PR 有99%的时间可以直接工作
50%UnverifiedDanny 用了六个月时间搭建 AgentOS，该系统接管了他约95%的编程和运营任务
50%UnverifiedAgentOS 中每个任务在独立的云容器中运行，任务完成后容器被销毁，下次会话需重新初始化项目
50%UnverifiedAgentOS 采用最小权限原则限制每个智能体能做什么，从基础设施层面限制模型输出可能造成的损害
50%Unverified在 AgentOS 中，客服机器人可访问 Front 客服系统的 MCP，但无法访问 Gmail 或 GitHub 仓库
50%UnverifiedAgentOS 中每个智能体的文件系统访问权限可精确到文件夹级别和操作类型级别（如只能写、不能删），由MCP服务端校验拦截违规操作
50%UnverifiedEwan is building an application called AgentOS on Agent Harness for assigning and tracking tasks for both himself and his agents.
50%