Building Effective Agents
by Anthropic
The essay that gave the field its vocabulary for workflows versus agents — and its case for building neither.
Overview
Published 19 December 2024 by two Anthropic engineers, Barry Zhang among them, this is the essay that gave the industry its shared vocabulary for agent design. Its central move is a definition: workflows are systems where LLMs and tools are orchestrated through predefined code paths, while agents are systems where the LLM dynamically directs its own process and tool usage. Everything follows from that split. The post starts from a single building block, the augmented LLM — a model with retrieval, tools and memory attached — then catalogues five workflow patterns: prompt chaining (sequential decomposition), routing (classify, then dispatch to a specialist), parallelisation (sectioning independent subtasks, or voting across repeated attempts), orchestrator-workers (a lead model that decomposes work at runtime) and evaluator-optimizer (a generate-then-critique loop). Only then does it describe true agents, which run tools in a loop against environmental feedback with explicit stopping conditions and human checkpoints. A section on frameworks — Claude Agent SDK, AWS's Strands Agents SDK, Rivet, Vellum — argues you should start with raw LLM API calls and adopt a framework only once you understand what it hides. Two appendices ground the theory: real customer-support and coding-agent deployments including SWE-bench Verified, and a guide to prompt-engineering the agent-computer interface. It is framework-neutral prose with linked cookbook implementations, not a tutorial, and it drew 763 points on Hacker News.
At a Glance
- Topic
- Agentic
- Level
- Intermediate
- Format
- Guide
- Cost
- Free
- Duration
- ~30 min read
- Provider
- Anthropic
- Hands-on
- No
- Certificate
- None
What You’ll Learn
- ✓A precise definition separating orchestrated workflows from genuinely autonomous agents
- ✓Five composable workflow patterns and the specific failure each one fixes
- ✓When added autonomy is worth its latency, cost and error-compounding penalty
- ✓How to structure an agent loop with feedback, stopping conditions and human checkpoints
- ✓Why to start with raw API calls before adopting any agent framework
- ✓How to prompt-engineer tool definitions — the agent-computer interface — for reliability
Highlights
- •Written by the engineers who ship Claude's own agent products
- •Framework-neutral: the patterns apply on LangGraph, the Agent SDK, or nothing at all
- •763 points and 121 comments on Hacker News, plus 543 more on the engineering re-post
- •Two appendices with real deployments, including SWE-bench Verified coding agents
- •Links to runnable cookbook implementations of every pattern it names
Who It’s For
Best For
- ✓Engineers choosing an architecture before writing a line of agent code
- ✓Tech leads deciding whether a task justifies an agent at all
- ✓Teams whose autonomous agent is unpredictable and expensive in production
Prerequisites
- •Working understanding of LLM prompting and tool or function calling
- •No code required — this is a design essay, not a tutorial
FAQ
What is Building Effective Agents?
Anthropic's canonical engineering essay on agent design, published December 2024 and still the reference text. It separates predefined-path workflows from autonomous agents, catalogues five composable workflow patterns, and argues for the simplest architecture that works. You finish with the vocabulary and the decision criteria to choose an agent architecture — and to recognise when you should not build an agent at all.
Is Building Effective Agents free?
Building Effective Agents is free to access.
What level is Building Effective Agents for?
Building Effective Agents is aimed at a intermediate audience. Recommended background: Working understanding of LLM prompting and tool or function calling, No code required — this is a design essay, not a tutorial.
How long does Building Effective Agents take?
Expect roughly ~30 min read. Most learners work through it at their own pace.
What will I learn from Building Effective Agents?
You'll learn: A precise definition separating orchestrated workflows from genuinely autonomous agents; Five composable workflow patterns and the specific failure each one fixes; When added autonomy is worth its latency, cost and error-compounding penalty; How to structure an agent loop with feedback, stopping conditions and human checkpoints; Why to start with raw API calls before adopting any agent framework; How to prompt-engineer tool definitions — the agent-computer interface — for reliability.
Topics
Sources
This page was written from 3 sources, 2 on domains other than anthropic.com.