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

Long Running​

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

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.

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​