Agentic Patterns

Source: https://github.com/nibzard/awesome-agentic-patterns at 87538fdeccf6
License: Apache-2.0
Read the English original

This edition is an AI-assisted translation.

Awesome Agentic Patterns

精選的代理式 AI 模式目錄，收錄協助自主或半自主 AI 代理在正式環境中完成工作的實用技巧、工作流程與小型架構。

為什麼？ 教學展示的是玩具範例，真實產品則隱藏了繁雜細節。本清單整理可重複使用的模式來彌合差距，讓大家都能推出更聰明、更快速的代理。

---

什麼才算是一種模式？

• 可重複——不只一個團隊正在使用。
• 以代理為中心——改善 AI 代理的感知、推理或行動方式。
• 可追溯——有公開資料支持，例如部落格文章、演講、程式碼庫或論文。

如果你的連結符合這些條件，就適合收錄在這裡。

---

🌐 探索網站

造訪： https://agentic-patterns.com

網站還提供 README 之外的強大探索工具：

• Pattern Explorer：依類別、狀態、複雜度等瀏覽、篩選及搜尋所有模式
• Compare Tool：並列比較多個模式的共同屬性
• Decision Explorer：互動指南，協助為使用情境找到合適模式
• Graph Visualization：以圖形呈現模式間的關係與連結
• Pattern Packs：為常見代理架構精選的模式集合
• Developer Guides：深入介紹模式選擇與使用方式的文件
• Dark Mode：完整主題支援，讓你在任何環境都能舒適閱讀

使用 Astro 建置，部署於 Vercel，原始碼位於 apps/web/。

---

類別快速導覽

類別	內容
情境與記憶	滑動視窗整理、向量快取、情節記憶
回饋迴圈	編譯器、CI、人工審查、自我修復重試
學習與調適	Agent RFT、技能庫、依變異數選樣的 RL
協調與控制	工作分解、產生子代理、工具路由
可靠性與評估	防護措施、評估框架、記錄、可重現性
安全與防護	隔離 VM、個人資料識別碼權杖化、安全掃描
工具使用與環境	Shell、瀏覽器、資料庫、Playwright、沙箱技巧
使用者體驗與協作	提示交接、暫存提交、非同步背景代理

類別並非一成不變——若你想到更合適的分類，歡迎提交 PR！ 下方表格由 patterns/ 資料夾自動產生。

---

上下文與記憶

• Agent-Powered Codebase Q&A / Onboarding
• Context Budget as a Governed Resource
• Context Window Anxiety Management
• Context Window Auto-Compaction
• Context-Minimization Pattern
• Cross-Agent Lesson Sharing via Git
• Curated Code Context Window
• Curated File Context Window
• Dynamic Context Injection
• Episodic Memory Retrieval & Injection
• Filesystem-Based Agent State
• Layered Configuration Context
• Memory Synthesis from Execution Logs
• Proactive Agent State Externalization
• Progressive Disclosure for Large Files
• Prompt Caching via Exact Prefix Preservation
• Schema-Guided Graph Retrieval for Multi-Hop Reasoning
• Self-Identity Accumulation
• Semantic Context Filtering Pattern
• Session-Scoped Context Runtime for Agent Tools
• Tool Search Lazy Loading
• Working Memory via TodoWrite

回饋循環

• AI-Assisted Code Review / Verification
• Background Agent with CI Feedback
• Coding Agent CI Feedback Loop
• Deterministic Grader in the Loop
• Dogfooding with Rapid Iteration for Agent Improvement
• Graph of Thoughts (GoT)
• Incident-to-Eval Synthesis
• Inference-Healed Code Review Reward
• Iterative Prompt & Skill Refinement
• Reflection Loop
• Rendered UI Finish Gate
• Rich Feedback Loops > Perfect Prompts
• Self-Critique Evaluator Loop
• Self-Discover: LLM Self-Composed Reasoning Structures
• Spec-As-Test Feedback Loop
• Tool Use Incentivization via Reward Shaping

學習與適應

• Agent Reinforcement Fine-Tuning (Agent RFT)
• Commitment Ledger with Reality-Gated Credit
• Compounding Engineering Pattern
• Frontier-Focused Development
• Memory Reinforcement Learning (MemRL)
• Persistent Test Memory Feedback Loop
• Shipping as Research
• Skill Library Evolution
• Variance-Based RL Sample Selection

編排與控制

• 402-First Machine Payments (Price-Before-Work Tool Purchases)
• Action-Selector Pattern
• Agent Modes by Model Personality
• Agent-Driven Research
• Artifact-Driven Analysis Pipeline Orchestration
• Autonomous Workflow Agent Architecture
• Board-Mediated Async Inter-Agent Coordination
• Budget-Aware Model Routing with Hard Cost Caps
• Burn the Boats
• Capability-Escrow-Receipt
• Classify-Then-Act for Background Agents
• Conditional Parallel Tool Execution
• Continuous Autonomous Task Loop Pattern
• Cross-Cycle Consensus Relay
• Cross-Domain Agent Conflict Resolution
• Custom Sandboxed Background Agent
• Declarative Multi-Agent Topology Definition
• Design-Time File Partition as Concurrency Control
• Deterministic Zero-LLM Orchestration
• Discrete Phase Separation
• Disposable Scaffolding Over Durable Features
• Distributed Execution with Cloud Workers
• Dual LLM Pattern
• Dual-Rail Message Delivery
• Economic Value Signaling in Multi-Agent Networks
• Explicit Posterior-Sampling Planner
• Factory over Assistant
• Feature List as Immutable Contract
• Hybrid LLM/Code Workflow Coordinator
• Inference-Time Scaling
• Initializer-Maintainer Dual Agent Architecture
• Inversion of Control
• Iterative Multi-Agent Brainstorming
• Lane-Based Execution Queueing
• Language Agent Tree Search (LATS)
• LLM Map-Reduce Pattern
• Markdown Polis — Multi-Vendor Agent Coordination via Filesystem Constitution
• Multi-Model Orchestration for Complex Edits
• Non-Generative Judgment Routing with Typed Escalation
• Opponent Processor / Multi-Agent Debate Pattern
• Oracle and Worker Multi-Model Approach
• Parallel Tool Call Learning
• Plan-Then-Execute Pattern
• Planner-Worker Separation for Long-Running Agents
• Progressive Autonomy with Model Evolution
• Progressive Complexity Escalation
• Recursive Best-of-N Delegation
• Self-Rewriting Meta-Prompt Loop
• Signal-Driven Agent Activation
• Specification-Driven Agent Development
• Stop Hook Auto-Continue Pattern
• Sub-Agent Spawning
• Subject Hygiene for Task Delegation
• Swarm Migration Pattern
• Three-Stage Perception Architecture
• Tool Capability Compartmentalization
• Tool Selection Guide
• Tracker-as-Desired-State Reconciliation
• Tree-of-Thought Reasoning
• Workspace-Native Multi-Agent Orchestration

可靠性與評估

• Action Caching & Replay Pattern
• Adaptive Sandbox Fan-Out Controller
• Agent Circuit Breaker
• Anti-Reward-Hacking Grader Design
• Asynchronous Coding Agent Pipeline
• Canary Rollout and Automatic Rollback for Agent Policy Changes
• CriticGPT-Style Code Review
• Dead-Man’s Switch for Scheduled Agent Jobs
• Evidence-Layered Evaluation for Interactive Agents
• Extended Coherence Work Sessions
• Failover-Aware Model Fallback
• Lethal Trifecta Threat Model
• LLM Observability
• Merged Code + Language Skill Model
• No-Token-Limit Magic
• Orchestration Prompt-Writing Benchmark
• Out-of-Process Provider-Boundary Replay
• Output Verification Loop
• Own-Check Fault Injection
• Reasoning-Token Firewall
• Reliability Problem Map Checklist for RAG and Agents
• RLAIF (Reinforcement Learning from AI Feedback)
• Schema Validation Retry with Cross-Step Learning
• Skill Activation as a Precision/Recall Measurement
• Structured Output Specification
• Subagent Compilation Checker
• Tier Auto-Apply by Mechanical Impact
• Versioned Constitution Governance
• Workflow Evals with Mocked Tools

安全與防護

• Authenticated Authority Channel
• Black-Box Skill Invocation
• Consequence-Family Coverage Audit
• Cryptographic Governance Audit Trail
• Denial Tracking & Permission Escalation
• Deterministic Security Scanning Build Loop
• Deterministic Threat Rule Scanning
• Evidence Questions, Not Verdict Questions
• Exact-Action Authorization Binding
• External Credential Sync
• Hook-Based Safety Guard Rails for Autonomous Code Agents
• Isolated VM per RL Rollout
• Local-First Credential Broker
• Non-Custodial Spending Controls
• One OS User per Agent
• PII Tokenization
• Policy-Gated Tool Proxy
• Sandboxed Tool Authorization
• Soulbound Identity Verification
• Transitive Vouch-Chain Trust
• Zero-Knowledge Verified Agent Egress
• Zero-Trust Agent Mesh

工具使用與環境

• Agent SDK for Programmatic Control
• Agent-First Tool Discovery
• Agent-First Tooling and Logging
• Agentic Search Over Vector Embeddings
• AI Web Search Agent Loop
• CLI-First Skill Design
• CLI-Native Agent Orchestration
• Code Mode MCP Tool Interface Improvement Pattern
• Code-Over-API Pattern
• Code-Then-Execute Pattern
• Cross-Protocol Agent Discovery
• Dual-Use Tool Design
• Dynamic Code Injection (On-Demand File Fetch)
• Egress Lockdown (No-Exfiltration Channel)
• Filesystem-Mediated Host Delegation
• Intelligent Bash Tool Execution
• LLM-Friendly API Design
• MCP Pattern Injection
• Multi-Platform Communication Aggregation
• Multi-Platform Webhook Triggers
• Patch Steering via Prompted Tool Selection
• Precomputed Code Graph Lookup
• Progressive Tool Discovery
• Shell Command Contextualization
• Static Service Manifest for Agents
• Tool Use Steering via Prompting
• Unified Tool Gateway
• Virtual Machine Operator Agent
• Visual AI Multimodal Integration

使用者體驗與協作

• Abstracted Code Representation for Review
• Agent-Assisted Scaffolding
• Agent-Friendly Workflow Design
• AI-Accelerated Learning and Skill Development
• Chain-of-Thought Monitoring & Interruption
• Codebase Optimization for Agents
• Democratization of Tooling via Agents
• Dev Tooling Assumptions Reset
• Human-in-the-Loop Approval Framework
• Latent Demand Product Discovery
• Milestone Escrow for Agent Resource Funding
• Proactive Trigger Vocabulary
• Seamless Background-to-Foreground Handoff
• Spectrum of Control / Blended Initiative
• Team-Shared Agent Configuration as Code
• Verbose Reasoning Transparency

---

提供 AI 助理使用（llms.txt）

本專案包含 llms.txt，這是一份機器可讀文件，旨在協助 AI 助理和 LLM 理解並推薦合適的模式。

包含內容：

• 模式類別及其用途
• 附有簡要說明的重要模式
• AI 助理使用指南
• 依使用情境需求選擇模式的策略

給建置 AI 助理的開發者： 可將 llms.txt 檔案提供給 LLM 作為上下文，以改善模式推薦。此檔案特別適用於：

• 為此目錄建立索引的 RAG 系統
• 推薦模式的 AI 程式設計助理
• 推薦代理式模式的 LLM 工具

存取： https://agentic-patterns.com/llms.txt （也可在 apps/web/public/llms.txt 中取得）

---

三步參與貢獻

1. Fork 並建立分支 → git checkout -b add-my-pattern
2. 使用上方範本在 patterns/ 下新增模式檔案。
3. 執行 bun run build:data，更新 README 自動產生的章節及網站資料。
4. 開啟 PR，標題為 Add: my-pattern-name。
5. 本儲存庫以模式為核心：即使提案在技術上有效，若主要是產品公告或推廣，仍會遭到拒絕。

詳細規定請參閱 CONTRIBUTING.md。

---

靈感來源

本專案起源於文章 「Sourcegraph 在建置 AI 程式設計代理時的經驗」（2025 年 5 月 28 日）以及持續更新的 Raising an Agent 影片日誌。許多早期模式直接來自這些經驗——感謝所有公開分享歷程的人！

---

授權條款

Apache‑2.0。請參閱 LICENSE。

---

星星數歷史