Building Effective AI Agents

Anthropic 官方 agent 构建指南，workflows vs agents 分类，五种 workflow 模式，ACI 设计原则

源

SOURCE · BUILDING EFFECTIVE AGENTS · Schluntz & Zhang · Anthropic 2024

Building Effective AI Agents

Schluntz & Zhang (Anthropic, 2024-12-20) — lessons distilled from dozens of customer agent builds

Core claim: the most successful implementations don’t lean on heavy frameworks — they use simple, composable patterns. The base building block is the “augmented LLM”: retrieval + tools + memory, with standardized interfaces via protocols like MCP. A single LLM call + retrieval + in-context examples is usually enough — only add complexity when there’s evidence it improves outcomes.

Five workflows + one agent pattern

Prompt ChainingTask decomposed into sequential steps; gates optional between them

RoutingClassify input, then dispatch to specialized handlers — semantic routing

ParallelizationSectioning (split work) + voting (ensemble of answers)

Orchestrator-WorkersCentral LLM dynamically decomposes the task, dispatches to workers, synthesizes results

Evaluator-OptimizerGenerate-evaluate iteration — feedback-driven refinement

Agent (autonomous)LLM dynamically directs its own flow — open-ended problems with unpredictable step counts

ACI design principle

ACI matters as much as HCI

Write tool definitions with the clarity you’d use in docs for a junior developer — the agent is the primary user of the tool

→ Prompt Chaining · Orchestrator-Workers · Augmented LLM · ACIAnthropic (2024-12-20)

Building Effective AI Agents

来源: sources/anthropic_official/building-effective-agents.md
URL: https://www.anthropic.com/research/building-effective-agents
作者: Erik Schluntz, Barry Zhang (Anthropic)
发布: 2024-12-20

摘要

Anthropic 基于与数十个客户团队合作构建 LLM agent 的经验总结。核心论点：最成功的实现不依赖复杂框架，而是用简单、可组合的模式构建。文章区分了 workflows（预定义代码路径编排 LLM）和 agents（LLM 动态指挥自身流程），并系统梳理了五种 workflow 模式和一种 agent 模式。

关键要点

从简单开始。只在有证据表明复杂度能改善结果时才增加复杂度。单次 LLM 调用 + 检索 + 上下文示例通常已经够用。
基础构建块是增强型 LLM（augmented LLM）：检索、工具、记忆三项增强能力，通过 MCP 等协议标准化接口。
五种 workflow 模式：
- Prompt chaining — 任务分解为顺序步骤，中间可加检查门
- Routing — 输入分类后导向专门化处理
- Parallelization — 并行处理（sectioning 分段 + voting 投票）
- Orchestrator-workers — 中央 LLM 动态分解任务、分发、综合
- Evaluator-optimizer — 生成-评估循环迭代
Agent 适用场景：开放式问题、步骤数不可预测、需要自主决策。代价是更高成本和错误累积风险。
ACI（Agent-Computer Interface）设计与 HCI 同等重要。工具定义要像写给初级开发者的文档一样清晰。

与其他 source 的关联

本文是 Anthropic 官方 agent 架构模式总览，后续的 harness 系列文章深入讨论了长时运行场景下的 harness 设计。
Claude Agent SDK 是本文推荐的实现框架。

References

sources/anthropic_official/building-effective-agents.md