Agent frameworks

Frameworks give you the loop, tool schemas, state and tracing. Choose by how much control you need.

Changes fast: check the latest updates too. Reviewed 17 Oct 2026.

Key ideas

• Pydantic AI: typed, lean, good structured outputs and dependency injection.
• LangGraph: explicit state graphs, checkpoints and human-in-the-loop; good for complex workflows.
• Strands Agents (AWS): model-driven loop, integrates with Bedrock and AgentCore.
• OpenAI Agents SDK, Claude Agent SDK, Google ADK: vendor SDKs with built-in tools and tracing.
• LlamaIndex: strongest on data connectors and retrieval.
• A raw API loop of about 50 lines is often enough and easiest to debug.

Common mistakes

• Framework lock-in through deep abstractions; keep prompts and tools as plain code you own.

Go deeper

• Pydantic AI
• LangGraph
• Strands Agents