§ Recipe
Build an AI agent from scratch.
Most agent tutorials are toys. A real agent has memory, tool use, guardrails, and observability. This is the stack that survives week two, not just the launch tweet.
Reasoning engine
01Claude's tool-calling and long-context behavior is the most reliable substrate for agents that need to reason over documents and invoke tools in sequence without going off the rails.
Orchestration framework
02Batteries-included patterns for chains, memory, retries, and tool routing. You'll outgrow it — but it saves six weeks of scaffolding you don't want to write.
RAG layer
03If your agent needs to know your data, LlamaIndex is the fastest path from documents to a production retrieval pipeline. Better defaults than rolling your own.
Tool access
04Instead of writing OAuth flows for every third-party integration, Composio gives your agent 200+ pre-wired tools. You go from GitHub-Slack-Notion in a day.
Multi-agent coordination
05When one agent isn't enough. CrewAI's role-based coordination beats hand-rolled message passing when you need researcher-plus-writer-plus-reviewer workflows.
Inference for sub-tasks
06For the low-stakes reasoning inside the loop — classification, routing, summarization — Groq's LPU inference is 10x cheaper and 5x faster than frontier model calls. Use it everywhere it's good enough.