Loading video player...
Most engineers pick an AI agent framework by GitHub stars or Twitter momentum — then pay the migration tax when abstractions shift underneath them. This video introduces a four-axis evaluation model that survives product rebrands, version bumps, and vendor pivots. Instead of ranking frameworks, we map the 2026 landscape — LangGraph, CrewAI, Claude Agent SDK, OpenAI Agents SDK, Pydantic AI, AutoGen, AG2, Agent Development Kit, and Sim — onto four durable design axes: - **Orchestration vs. Autonomy**: Who decides the next step — your explicit control flow or the model's reasoning loop? LangGraph sits at the orchestration pole; Claude Agent SDK leans toward autonomy. - **Code vs. Config**: Is the agent defined in typed Python functions or YAML configuration files? Pydantic AI is strongly code-first; CrewAI recommends YAML for agent and crew definitions. - **Single vs. Multi-Agent**: Do you start with one capable agent or a team of specialists? Most practitioners — including Anthropic's own research guidance — recommend starting single and scaling out only when the workload demands it. - **Deterministic vs. Reasoning**: Where do you draw hard boundaries around the model? Every production system mixes both; the question is where the framework makes each easy or painful. We apply the axes to a concrete Email-to-Action Agent workflow, showing how real constraints (draft approval, sender history, outbound formatting) calculate a specific coordinate that points to a framework — before any leaderboard enters the conversation. The video also covers the migration tax (including Microsoft's Semantic Kernel → Agent Framework shift in 2026), why GitHub stars and PyPI downloads tell contradictory stories, and what experienced practitioners like Hamel Husain actually optimize for. Part of the Human Architects series. For the full research and long-form arguments, see the companion book "AI Agents in Action." #agentframeworks #llm #langgraph #crewai #pydanticai #autogen #aiagents #softwarearchitecture 📑 Chapters: 0:00 The migration tax: what choosing wrong actually costs 0:44 Why GitHub stars and PyPI downloads contradict each other 1:46 The four durable axes of agent architecture 2:00 Axis 1: Orchestration vs. autonomy 2:37 Axis 2: Code vs. configuration 3:20 Axis 3: Single-agent vs. multi-agent 4:05 Mapping real constraints to the coordinate grid 4:50 Email-to-Action Agent: applying the axes 5:34 What experienced practitioners actually optimize for 6:18 Operational axes: durability, observability, governance 6:45 When the map signals a framework shift 7:12 Pick the position, not the logo #agent frameworks #langgraph #crewai #pydantic ai #claude agent sdk #openai agents sdk #autogen #ag2 #ai agents #agent architecture #framework comparison #software architecture #migration tax #orchestration vs autonomy