A 3,500-word pillar covering LangGraph, CrewAI, AutoGen, memory, eval, and cost. This pillar goes deeper than a summary — expect production code, numbers from real deployments, and opinionated trade-offs.
Most teams building “AI agents” in 2026 aren't. They're building LLM-augmented chatbots — a single call per user turn, some RAG, maybe a tool or two. That's fine, but it's not an agent.
1. Topology: single-graph vs hierarchical
The choice of topology is the single most consequential architectural decision. A flat state machine is simpler to debug but caps out around 8–12 nodes.
2. Memory: episodic vs semantic vs working
The three memory types compose; they don't compete. A well-engineered agent uses all three, with distinct retrieval policies and retention windows.
Most teams under-invest in working memory — the scratchpad the agent uses within a single task — and massively over-invest in vector-search “long-term memory” that ends up polluting context.