AGENT PRODUCTION ENGINEERING
BUILD AGENTS
THAT SURVIVE
PRODUCTION.
Production-grade engineering patterns for autonomous AI agents.
让 AI Agent 从「能运行」，进化到「能够长期稳定运行」。
DIAGNOSIS
WHY AGENTS FAIL IN PRODUCTION
Agent 在生产环境里的失败不是随机的，而是六类结构性问题。每一条都对应一个已被验证的工程解法。
STATE LOSS
SESSION ENDS. THE AGENT FORGETS EVERYTHING.
会话结束，Agent 忘记一切。
ENGINEERING ANSWER→ STATE FILESILENT FAILURE
THE AGENT FAILED. NOBODY KNOWS.
任务失败了，没有人知道。
ENGINEERING ANSWER→ MAKER / CHECKERCONTEXT EXPLOSION
ERRORS KEEP ACCUMULATING. CONTEXT KEEPS GROWING.
错误不断堆积，上下文持续膨胀。
ENGINEERING ANSWER→ ERROR COMPACTSELF VALIDATION
THE AGENT VALIDATES ITSELF. THE VALIDATION IS NOT INDEPENDENT.
自己给自己打分，验证并不独立。
ENGINEERING ANSWER→ MAKER / CHECKERSKILL REGRESSION
NEW SKILL. OLD CAPABILITY BREAKS.
新模式上线，旧能力回归了。
ENGINEERING ANSWER→ EVOLUTION GATEPROBLEM → PATTERN → RESULT
EACH PROBLEM HAS AN ENGINEERING ANSWER
不是技巧清单，而是「问题 → 模式 → 结果」的因果链。
THE SYSTEM
ONE PRODUCTION-READY LOOP
调度触发 → 生成 → 独立验证 → 持久化 → 通知；失败则压缩错误重试，超限升级人工。节点可点击进入对应 Pattern。
这条流水线本身就是 Pattern 的组合：Cron Job 负责可靠触发，Maker / Checker 负责输出质量，State File + Checkpoint 负责断电恢复，Error Compact 负责失败不污染上下文。完整拆解见 Architecture 页。
PATTERNS WORK TOGETHER.
A SYSTEM, NOT A TRICK LIST
Production agents are built from systems of patterns, not isolated techniques. 悬停高亮关联，点击进入公约。
PATTERN LIBRARY
FEATURED PATTERNS
三个最核心的生产组件，建议按此顺序理解整个系统。
QUALITY
Maker/Checker 双角色分离
Separate generation from validation.
生成与验证不在同一个 Agent
STATE
STATE.md 跨运行状态管理
Cross-run memory: read before run, write after every step.
Read Before Run, Write After Every Step
AUTOMATION
Cron 任务设计模式
Idempotent, dedup-safe, silent-failure-proof scheduling.
幂等、防重复、防静默失败（含 Hermes 原生 Monitor 模式）
SECONDARY PATTERNS
MATURITY MODEL
FROM AGENT TO PRODUCTION SYSTEM
从 Prompt 到自主系统的六级阶梯；每一级由对应 Pattern 托底。
CASE STUDIES
RUNNING IN PRODUCTION, NOT IN A DEMO
四条 Pattern 在真实 7×24 工作流中的组合应用。
CRON SAFETY INTEGRATION
Wrap a scheduled Hermes task with maturity staging, state tracking, idempotency, compact errors, and hard stops.
→ VIEW CASE STUDYDAILY NEWS DIGEST
Run a daily digest loop that collects candidate links, filters duplicates, drafts a concise summary,
→ VIEW CASE STUDY实战案例S
本目录存放 Hermes Production Patterns 的真实落地案例。每个案例都遵循「先有真实运行，再有文档」——能复现的输出比描述更有说服力。
→ VIEW CASE STUDYMAKER/CHECKER ARTICLE PIPELINE
Maker 写稿后，启动一个独立的 Agent 实例加载 maker-checker skill 做验证：
→ VIEW CASE STUDYWECHAT ARTICLE PIPELINE
Use Hermes Agent to write, detect AI-taste, de-AI-fy, and publish WeChat Official Account articles.
→ VIEW CASE STUDYON-RAMP
CHOOSE YOUR PATH
按你的角色直接开始，30 秒找到第一个要读的 Pattern。
HOW READY IS YOUR AGENT?
PRODUCTION READINESS CHECK
回答四个问题，评估你的 Agent 的生产就绪程度。
Does your agent persist state across sessions?
START TODAY
BUILD AGENTS THAT SURVIVE PRODUCTION.
clone 仓库即可把全部 15 条公约装进你的 Hermes。