Framework Explorer
What each framework solves, how it is shaped, where it is weak, and when not to use it. Option 0 — no framework — comes first on purpose.
Nothing extra — the provider SDK already gives you messages, tool definitions, and structured output; a 30-line loop in your own code is the entire agent runtime.
A common interface over many LLM providers, vector stores, document loaders, and tools, so a pipeline can be composed from interchangeable parts.
Durable, resumable agent workflows expressed as an explicit state graph with checkpoints, so long-running and human-interrupted runs can pause and continue.
Fast assembly of role-based multi-agent teams where each agent has a persona, goal, and tools, and tasks are distributed sequentially or by a manager.
Ingesting, indexing, and querying private documents with LLMs — RAG as a first-class product rather than a chain you assemble.
A small, opinionated agent runtime with handoffs between agents, guardrails, and built-in tracing, tightly integrated with OpenAI models and hosted tools.
Type-safe agents where dependencies, tool arguments, and results are all Pydantic-validated, with the framework staying close to the underlying SDKs.
A provider-agnostic TypeScript layer for text generation, streaming, tool calling, and structured output, with React/Next.js hooks for streaming UIs.
Composable agent hierarchies (LLM agents plus deterministic sequential/parallel/loop agents) with a built-in dev UI and a path to Vertex AI Agent Engine deployment.
A single Microsoft-supported runtime for agents and multi-agent workflows in .NET and Python, merging Semantic Kernel's enterprise plumbing with AutoGen's multi-agent orchestration patterns.