The Best Articles on Building Agents
Will update this as I find more.
Building effective agents: Anthropic says the most reliable agents come from simple, composable patterns with clear tool interfaces—only add agentic complexity when measurements show it helps.
How we built our multi-agent research system: Anthropic built a multi-agent system where a lead agent coordinates parallel subagents to independently research subquestions, improving factual accuracy and coverage—but at the cost of high token usage, latency, and complex orchestration.
How to Build an Agent: Thorsten Ball demonstrates a code-editing agent in <400 lines by looping an LLM with a few filesystem tools and letting it decide when to call them—no fancy framework required.
Context Engineering for AI Agents: Lessons from Building Manus: Manus prioritizes “context engineering” (stable KV-cache prefixes, append-only deterministic traces, masking tools instead of removing them) to keep long-running agents fast, cheap, and robust.
Agentic Search for Dummies: Benjamin Anderson’s baseline is full-text search + offline augmentation, a Tantivy index, and simple search/read tools so the model can spray many queries and fuse results before deciding what to read.
Don’t Build Multi-Agents: Cognition’s Walden Yan argues that multi-agent architectures are currently brittle and overcomplicated, and instead recommends single-agent systems with tool-use and memory, emphasizing that action traces already encode planning.