AI Agent Patterns

Core Philosophy

Start simple, add complexity only when measured — Begin with a single ReAct loop. Add planning, reflection, or multi-agent orchestration only when evaluation shows the simpler pattern failing. Anthropic's #1 finding: the most successful agent builders resist unnecessary complexity.
Tools over prompts — Invest more time in tool definitions than prompt engineering. Well-specified tools with clear schemas, examples, and error messages outperform clever prompts. The tool is the agent's interface to the world.
Architecture determines cost, reliability, and scaling — A ReAct agent makes 5-7 LLM calls per task; Plan-and-Execute often cuts this to 3-4. Wrong architecture choice compounds across thousands of requests. Choose based on workload, not hype.
Memory is not one thing — Agents need multiple memory systems: sliding window for conversation, working memory for current task state, semantic memory for long-term knowledge, episodic memory for past experiences. Each serves a different retrieval need.
Evaluate before shipping — Agent performance drops 58% between single execution and eight consecutive runs. Without evaluation infrastructure, model upgrades take weeks instead of days. Build evals first.