An agent framework, in the classification Agent Frameworks, Runtimes, and Harnesses- oh my! proposes, is a package for building with LLMs whose main value add is abstractions. Harrison Chase, who wrote that post, describes those abstractions as representing a mental model of the world: ideally they make it easier to get started, and they give developers a standard way to build applications, so that onboarding and moving between projects is easier. He classifies most packages for building with LLMs this way, naming LangChain as LangChain’s own example alongside Vercel’s AI SDK, CrewAI, the OpenAI Agents SDK, Google’s Agent Development Kit (ADK) and LlamaIndex.
Usage
The term is defined by contrast with two neighbours in the same post. An Agent Runtime sits below a framework, supplying production infrastructure such as durable execution, and can power one — LangChain 1.0 is built on LangGraph. An Agent Harness sits above, adding default prompts, opinionated tool-call handling, planning tools and filesystem access on top of a framework. Chase stresses that the boundaries are blurry and that no clear definition yet exists; LangGraph, for example, is called both a runtime and a framework.
The standing criticism of the category is also about abstractions. Chase records the complaint that poorly made abstractions obscure how things work and do not give the flexibility advanced use cases need. A second LangChain post, Agent Middleware, states the same problem more specifically: frameworks built around the core loop of a model, a prompt and a list of tools do not give developers enough control over context engineering, so developers end up “graduating off of the abstraction” for any non-trivial use case. That post’s answer is to keep the loop and make it modifiable through Agent Middleware.
Related Terms
- Agent Runtime — the lower-level infrastructure layer
- Agent Harness — the higher-level, batteries-included layer
- Agent Middleware — LangChain 1.0’s abstraction for customizing a framework’s agent loop
- Context engineering — the control the middleware post says frameworks fail to give