LangChain
Agent-framework vendor and the fourth agent-platform maker tracked here (after anthropic,
google, microsoft). Its position: agents are model + harness, and a harness is best
built from composable agent-middleware, not a monolithic framework
langchain-custom-harness.
What it ships
create_agent— its minimal agent-loop builder, “the easiest way to build a custom agentic-coding-harness”; middleware stacks on top.- Deep Agents — a production-ready harness built on
create_agent(opinionated default stack). - A library of prebuilt middleware (summarization, context editing, memory, subagents, retries, PII, human-in-the-loop, prompt caching).
- Says its own GTM and coding agents run on
create_agent+ tailored middleware.
Positioning
Where google (adk) and microsoft (Agent Framework) lead with skills as the primitive (agentskills-spec), LangChain leads with middleware as the primitive — a different decomposition of the same harness. Both reject the monolith; the question is whether the reusable unit is a loadable skill or a composable middleware (often, both). It also leans on anthropic‘s Claude Agent SDK as a comparison point (a pre-assembled middleware harness).
Related
langchain-custom-harness · agent-middleware · agentic-coding-harness · anthropic · google · microsoft