Skip to main content

Patterns

Agent design patterns:

  • Give agents a computer (CLI and files)
  • Progressive disclosure
  • Offload context
  • Cache context
  • Isolate context
  • Evolve context

Agent Paradigm

Architecture​

General agent components:

  • LLM (brain)
  • Prompting (instructions)
  • Memory
  • External knowledge
  • Tools

AI Agent system use LLMs as core reasoning engine, augmented with tools, memory, and instructions.

Agent Architecture

Agent Loop

First-Principles​

从李世石与 AlphaGo 的围棋对战中的第 37 手, 我们可以总结出第一性原理 智能体的基本原则:

  • Replica agents: 当流程需要人工审核、代理作为用户的副驾驶员或与仅限 UI 的旧版工具集成时，使用仿生学。
  • Alien agents: 当目标是纯粹的结果效率时，使用第一性原理。

Asymmetry of Verification and Verifiers​

Asymmetry of verification and verifiers law:

所有可解决且易于验证的问题, 都将被 AI 解决.

Agent Traffic

Among the agents:

Value of highly polished UI and enterprise applications will decrease, value of performant, reliable, extensible API will increase.

Agent-Native​

Agent-native apps should:

  • Parity (对等性): 用户通过 UI 完成任务 <-> Agent 通过工具实现.
  • Granularity (细粒度): tools should be atomic primitives.
  • Composability: 有了上述两点, 只需编写新的提示词即可创建新功能.
  • Emergent capability.
  • Files as universal interface: files for legibility, databases for structure.
  • Improvement over time:
    • Accumulated context: state persists across sessions.
    • Developer-level refinement: system prompts.
    • User-level customization: user prompts.
**Who I Am**:
Reading assistant for the Every app.

**What I Know About This User**:
- Interested in military history and Russian literature
- Prefers concise analysis
- Currently reading *War and Peace*

**What Exists**:
- 12 notes in /notes
- three active projects
- User preferences at /preferences.md

**Recent Activity**:
- User created "Project kickoff" (two hours ago)
- Analyzed passage about Austerlitz (yesterday)

**My Guidelines**:
- Don't spoil books they're reading
- Use their interests to personalize insights

**Current State**:
- No pending tasks
- Last sync: 10 minutes ago
Agent-native Product

Build capable foundation, observe what users ask agent to do, formalize patterns that emerge:

  • Common patterns: domain tools.
  • Frequent requests: dedicated prompts.
  • Unused tools: remove.

Long Running​

Three-agent architecture harness design (Planner + Generator + Evaluator) produce rich full-stack applications over multi-hour autonomous coding sessions.

自主编码正从更好的提示转向更好的控制系统: 围绕 agent 设计目标、评估器、循环、产物, 使其在人停止输入后仍能规划、执行、自检、纠错、持续推进, agent 仍是执行者, 但人不再逐轮交互.

Loop Engineering

Goal as Contract​

/goal 规定期望终态、成功证据、约束、轮次与预算, 而非更长的提示:

  • 弱目标让模型提前停止或重定义成功
  • 强目标编码领域知识 (基准、截图、约束) 供反复自测

Verifier​

自主性只在拥有可靠验证器时成立:

  • 外部检查优于 agent 对已完成的自我解释
  • 确定性检查 (类型、测试、lint) 作下限, LLM 评审作高层复核
  • 详见 evaluation

Loop​

目标给方向, 循环让工作存活, 模型常在真正完成前停止:

  • 外层控制系统: 唤醒 -> 检查 -> 验证 -> 对照目标 -> 带回下一步
  • 最简形式即 Ralph loop + 确定性条件

Model Choice as Architecture​

模型是架构决策而非单一选择:

  • 规划模型定目标/约束, 执行模型跑实现, 廉价模型做评估/视觉评审
  • 编排器让你交换角色, 而非等待单一厂商

Artifacts​

多 agent 并行时终端记录不可扩展, 分离存储与呈现:

  • Markdown 存持久证据, 供 agent 搜索
  • HTML 产物渲染可视化仪表盘 (loss 曲线、基准、截图), 供人监控
  • 产物是控制面, 非事后报告

Session Mining​

过往会话是工作流数据富矿:

  • 重复失败、漏跑检查、重试坏命令不该埋在日志里
  • 扫描近期记录, 把重复失败模式转为项目指令/规则, 让本地环境更聪明而非从零训练模型

Self-Evolving​

Self-evolving agents, use runtime experience and external signals to optimize future behavior:

  • Update evaluation datasets.
  • Enhanced context engineering.
  • Tool optimization and creation.
  • Refine guardrails.

Flywheel

Compound​

Compound engineering (复利工程), 每个 PR 都在教育系统, 每个 bug 都成为永久的教训, 每次代码审查都在更新 agent 的默认行为:

  • 将经验沉淀到项目文档.
  • 让 bug 修复产生长期价值.
  • 从代码审查中提取模式.
  • 建立可复用的工作流程: slash commands, hooks, guardrails, and skills.
  • Linter rules, regression tests, AGENTS.md improvements, checklist updates.

AgentOps​

AgentOps

References​