The Agent Framework Landscape
A crowded framework market with one durable core — choose for typed contracts and escape hatches, not loyalty.
The agent-framework space is crowded: LangChain and LangGraph offer graph-orchestrated state machines, LlamaIndex centers retrieval-heavy workflows, CrewAI and AutoGen target role-playing multi-agent setups, while OpenAI Agents SDK, PydanticAI, Google ADK, and Anthropic Agent SDK compete on typed, lightweight primitives. Most share the same core of models, tools, state, and loops, so skills transfer. Choose by fit, not fashion: typed contracts and testability for finance work, escape hatches to raw APIs, and no lock-in to a vendor model. Frameworks here are scaffolding for the retrieval and multi-agent patterns that follow; the durable investment is understanding the loop, which survives any library rewrite.