Building Effective Agents

Overview

This article from Anthropic offers a practical, opinionated guide to building LLM-powered agents and workflows. Rather than jumping straight to complex autonomous agents, the authors argue for starting simple — using single LLM calls with retrieval and in-context examples before graduating to more complex agentic systems.

Key Workflow Patterns

The article defines a clear taxonomy of agentic systems, distinguishing between workflows (predetermined orchestration of LLMs) and agents (LLMs dynamically directing their own processes):

• Prompt Chaining — break a task into sequential steps, where each LLM call processes the output of the previous one. Useful when you can decompose a task cleanly.
• Routing — classify the input first, then direct it to a specialized handler. Good for complex tasks with distinct categories.
• Parallelization — run multiple LLM calls simultaneously and aggregate results. Works for tasks requiring multiple independent perspectives (e.g., voting on code quality).
• Orchestrator-Workers — a central LLM dynamically delegates subtasks to worker LLMs. Great when you cannot predict the subtasks in advance.
• Evaluator-Optimizer — one LLM generates, another evaluates, looping until quality criteria are met. Useful when you have clear evaluation standards.

When to Use Agents

The article recommends true agents (with tool use and autonomous loop control) only when the task requires flexibility and model-driven decision-making that workflows cannot handle. Agents trade latency and cost for better performance on open-ended problems.

Key Takeaways

1. Keep agentic systems as simple as possible — complexity for its own sake hurts reliability
2. Transparency is critical: show the agent’s planning steps to users
3. Invest heavily in your tool design — clear names, thorough descriptions, and well-tested interfaces make or break agent performance