/perspective/frameworks/
Filter the whole landscape down by the layer it owns and the patterns it makes native. Each card carries what it's for, the scenario it shines in, and how it talks to the wider ecosystem — then add two or three to a side-by-side comparison.
Where a framework sits is the first thing that separates it. Each layer answers a different question, governs different pattern domains, and stacks on top of the one below.
“How do I get started quickly?”
A shared vocabulary — agent loops, content blocks, model & tool integrations, middleware — without forcing you to learn the runtime underneath.
Key insightYou don't need to know the runtime to use a framework: LangChain runs on LangGraph internally, but you can build without ever touching it.
“How do I run this in production?”
The production infrastructure frameworks deliberately leave out: durable execution, streaming, human-in-the-loop, persistence and checkpoint-based recovery.
Key insightNeed fine-grained control, checkpoint-based recovery or explicit state? Build directly on a runtime rather than on a framework that sits on top of it.
“How do I handle complex, long-running tasks?”
An opinionated, batteries-included system for autonomous agents that decompose their own tasks and manage context over long runs. Sits on top of a runtime.
Key insightReach for a harness when you want an agent that decomposes its own tasks and manages context over long runs — not when you want to assemble the workflow yourself.
Event-driven conversation between agents; asynchronous messages and group chats.
Model-driven minimal SDK; the LLM weaves the graph from a 'bundle of threads' of tools.
Role-based 'Crews' of agents; flat abstraction over team-shaped workflows.
Modular, hierarchical workflow agents for the Vertex AI ecosystem.
Brings LLM and agentic capabilities to the JVM (Java/Kotlin) with LangChain-style abstractions.
Retrieval-centric framework; connectors and indexes turn proprietary documents into agent context.
SOP-based multi-agent metaprogramming that simulates an IT company.
Lightweight, Python-first SDK for handoff-based agent flows.
Type-safe Python agent framework; structured outputs and validation/retry loops on Pydantic v2.
Enterprise SDK for .NET/Java/Python; skills and planners under Microsoft/Azure governance.
Single-agent assistant platform for customer-facing deployments.
Event-driven durable execution; steps, retries and flow control over a queue you don't manage.
Graph-based, state-driven runtime; explicit nodes and edges, checkpointer-backed.
Durable execution for distributed services; durable promises, virtual objects and exactly-once handlers.
Durable workflow engine for minutes-to-hours processes; deterministic replay and automatic retries.
Anthropic's agentic harness — the gather-context / act / verify loop behind Claude Code, with tools, subagents and MCP.
Opinionated, batteries-included harness on LangGraph: planning, subagents and a virtual file system out of the box.
Where to next
Search patterns, frameworks, and pages.