Best practices for building agentic systems

A comprehensive look at best practices for building enterprise-grade agentic AI systems, drawing on insights from engineers at Shopify, Block, ServiceNow, Sonar, and others. Covers the core architectural components — reasoning models, context management, tool integration via MCP, multi-agent orchestration with LangGraph/CrewAI, security with just-in-time authorization, human-in-the-loop checkpoints, evaluation, and observability. Key guidance includes starting with specialist agents over generalists, avoiding multi-agent architectures early, using progressive context disclosure, codifying deterministic workflows, and embedding guardrails in IAM policies rather than prompts. The article also looks ahead to agent factories, edge inference, and the A2A protocol.