Agent loops & harness engineering,
small enough to read in an evening.
Agent loop 与 harness engineering，
小到一个晚上就能读完。
Every idea that reshapes AI engineering eventually gets a minimal implementation you can hold in your head. The nano series is that for agent loops and harness engineering — real, working systems, kept deliberately small, and mapped chapter-by-chapter onto the harness-101 course.
每个重塑 AI 工程的思想，最终都会有一个能装进脑子里的最小实现。nano 系列就是 agent loop 与 harness engineering 的最小实现——真实可运行的系统，刻意保持小巧， 并与 harness-101 课程逐章对应。
01The projects系列项目
Three layers of one stack: a single agent loop, an orchestrator that runs many of them, and the distributed backend that runs the fleet.
同一套技术栈的三个层次：单个 agent loop、批量调度它的编排器、以及运行整个集群的分布式后端。
A minimal AI coding assistant — the agent loop, tools, and context management without framework bloat. TUI + daemon modes, swarm multi-agent, sandboxed by default.
极简 AI 编程助手——agent loop、工具与上下文管理，没有框架包袱。TUI + daemon 双模式，swarm 多智能体，默认沙箱隔离。
A lightweight orchestration service for running coding agents on tracked issues — plan runs with DAG execution, human approval gates, cost budgets, and a web dashboard.
轻量级 agent 编排服务——把 coding agent 跑在可追踪的 issue 上：DAG plan run、人工审批门禁、成本预算和 Web 仪表盘。
Distributed execution for agents — a gateway/worker architecture that runs agent fleets in isolated Docker runtimes, with pairing-based onboarding and an admin console.
agent 的分布式执行——gateway/worker 架构，在隔离的 Docker 运行时里运行 agent 集群，配对式接入和管理控制台。
02One methodology, two ways in一套方法论，两种打开方式
03The 12 factors, mapped to running code12 因素 → 运行中的代码
Each factor links to the course chapter, the starter code, and the nano implementation.
每个因素都链到课程章节、骨架代码和 nano 中的实现。
04A suggested path建议的学习路径
- Work through the course prep and F1–F3 while reading nano-agent's core:
pkg/agentis the loop the course asks you to write, at production scale. - 完成课程准备和 F1–F3，同时对照读 nano-agent 的核心：
pkg/agent就是课程让你手写的那个 loop 的生产级版本。 - F4–F6: termination, failure, and permissions — study
pkg/agent/turn_policy.goto see how nano-agent decides a run is done (or stuck), then run it in its strictest permission mode and watch it refuse things. - F4–F6：终止判定、失败处理与权限模型——读
pkg/agent/turn_policy.go看 nano-agent 如何判定任务完成（或陷入循环），然后把权限开到最严，看它如何拒绝操作。 - F7–F9: graduate to nano-symphony — approval gates, plan runs, and isolated workspaces with durable artifacts are orchestrator concerns, not agent concerns.
- F7–F9：进阶到 nano-symphony——审批门禁、plan run、隔离工作区与持久 artifacts 是编排层的关注点，而不是 agent 本身。
- F10–F12: mailbox, swarm, and evaluation — then the capstone: deploy your own fleet on nano-cloud.
- F10–F12：mailbox、swarm 与评估——最后是毕业项目：在 nano-cloud 上部署你自己的 agent 集群。