LangChain's open source agent stack now has three layers: LangGraph (agent runtime), LangChain (agent framework), and Deep Agents (agent harness). LangGraph offers maximum determinism via graph-based workflows with human-in-the-loop and fault tolerance. LangChain provides a minimal, un-opinionated agent loop with middleware hooks for customization. Deep Agents, launched in July 2025 and inspired by Claude Code and Manus, bundles context management best practices including filesystem, subagents, skills, and memory out of the box. All three are composable and built on the same underlying philosophy. The recommended starting point is Deep Agents via `create_deep_agent`, dropping down to LangChain or LangGraph only when more control or determinism is needed.
Table of contents
What each layer offersWhen to reach for eachBalancing determinism and agencyWhy three layersTL;DRQuestions this post answers
What is the difference between Deep Agents, LangChain, and LangGraph?
Deep Agents is an opinionated agent harness with built-in context management (filesystem, subagents, skills, memory). LangChain is a minimal, un-opinionated agent framework — an LLM running in a loop calling tools, extensible via middleware. LangGraph is the underlying graph-based runtime offering maximum determinism, human-in-the-loop, fault tolerance, and observability. All three are composable; LangGraph powers the agent abstractions in both LangChain and Deep Agents. Teams choosing between these layers for production agents track the evolving tradeoffs on daily.dev.
When should I use LangGraph instead of LangChain or Deep Agents?
Use LangGraph when your workflow doesn't fit a standard agent loop or you need to mix deterministic and agentic steps. A good example is a rental application pipeline where only one step uses an LLM for extraction while the rest is fixed code. LangGraph lets you encode domain logic directly into the graph's topology rather than leaving decisions to the model. Developers designing custom agent workflows find architecture discussions like this on daily.dev.
When was Deep Agents released and what inspired it?
Deep Agents launched in July 2025, built on LangChain's core agent loop but with context management components — filesystem, subagents, skills, and memory — bundled by default. It was inspired by Claude Code and Manus, with the goal of giving builders equally powerful general-purpose agents for their own use cases via the `create_deep_agent` API. Keeping up with new agent framework releases like Deep Agents is easier when the news comes to you on daily.dev.