LangChain is the largest open source LLM application framework, repositioned in 2026 as an “agent engineering platform” spanning the create_agent harness, LangGraph orchestration, Deep Agents, and a terminal coding agent called dcode.
The coding-agent angle is now real and first-party: LangChain ships its own terminal coding agent, which moves it from “tooling you might build on” to “competitor in your harness choice”.
What it is #
The MIT-licensed OSS stack has three layers: LangChain’s create_agent (a minimal model plus tools plus middleware harness), LangGraph (low-level durable agent workflows), and Deep Agents (batteries-included planning, subagents, virtual filesystem, context compression).
Deep Agents Code (dcode) is an open source terminal coding agent with persistent memory, skills, subagents, remote sandboxes, and approval gates, installable independently of any LangChain service.
The commercial side is LangSmith: tracing, evaluation, deployment, sandboxes, and the no-code Fleet agents.
Status #
Active and dominant by footprint.
The langchain-ai/langchain repository shows 147.4k stars and 24.7k forks as of 2026-10-03, with 16,914 commits as of 2026-10-03.
The telling history: after the 2024 “death by abstraction” wave, the company publicly moved to lower-level primitives (LangGraph, then create_agent), and is now climbing back up with Deep Agents, dcode, and OpenWiki, a CLI that writes agent wikis for coding agents.
Strengths #
- The most portable model interface in the field: one API across OpenAI, Anthropic, Google, Bedrock, Ollama, and dozens more, which is the layer that actually resists churn.
- Retrieval docs now teach agentic RAG (2-step, agentic, hybrid) instead of pushing one big retrieval chain.
- dcode inherits Deep Agents’ context compression and subagents, and runs with any provider key you own.
- LangSmith evaluation and tracing remain the most mature observability pair for agents built this way.
Cautions #
- Trust debt is the core risk: the Octomind post (480 points, 297 comments) documents teams abandoning the framework for leaky abstractions, and while the CEO responded by moving investment lower-level, every new layer (Deep Agents, dcode) asks users to be burned twice.
- Ecosystem sprawl: choosing between LangChain, LangGraph, and Deep Agents is genuine design work, and the docs themselves need a decision tree for it.
- The retrieval stack is generic text RAG: the current text splitters catalog offers separator-based code splitting only, with no AST chunker.
- The polished experience assumes LangSmith, a paid SaaS, for tracing and evaluation.
Pricing #
OSS (LangChain, LangGraph, Deep Agents, dcode): free, MIT. LangSmith as of 2026-10-02: Developer $0 with 5k base traces/month, Plus $39/seat/month with 10k base traces, Enterprise custom, with usage metered in LangChain Standard Units at $1.00 per LSU across Deployment, Sandboxes, Engine, and Fleet. The framework is free forever; the operations layer around it is where the bill lives.
Price history #
| Date | Plan | Change | Source |
|---|---|---|---|
| 2026-09-18 | LangSmith | Baseline: OSS free (MIT); LangSmith Developer $0 (5k traces/mo), Plus $39/seat/mo (10k traces), Enterprise custom, metered units LCU $1.50 and LSU $1.00. | |
| 2026-10-02 | LangSmith | The LCU ($1.50) line item no longer appears on the pricing page; usage metering is consolidated in LSUs at $1.00, plan prices unchanged. |
Compared to #
- LlamaIndex: deeper retrieval modules, thinner agent platform; pair them or split by primary need.
- Established terminal harnesses (Claude Code, Codex, OpenCode): dcode is a credible new entrant but the incumbents have years of editor, CI, and workflow integration.
- Hand-rolled agent loops: for a single provider and a stable tool set, direct SDK plus a loop avoids the abstraction tax the HN thread documents.
Bottom line #
Recommended for teams building multi-provider agent systems who will also adopt LangSmith, and for anyone wanting an ownable coding agent they can fork rather than subscribe to. I would not pick LangChain in 2026 purely for RAG, its original job: retrieval has commoditized into provider APIs and agentic search, and the differentiating value has moved to the harness and evaluation layers. Disagree if you like; the retrieval-first framing is my call, not the docs'.
Changes #
- 2026-08-24 - Created among the seed notes of the Retrieval category.
- 2026-09-18 - Refreshed volatile facts for the 2026-09-18 verification: 146.6k stars, 24.5k forks, 16,782 commits, LangSmith pricing re-confirmed unchanged.
- 2026-09-20 - Added the Price history section tracking price changes in a table, per the new owner rule.
- 2026-09-25 - Refreshed volatile facts: 147.0k stars, 24.6k forks, 16,852 commits as of 2026-09-25; LangSmith pricing re-confirmed unchanged.
- 2026-10-02 - Recorded the LangSmith pricing page consolidating metering in LSUs at $1.00 with the LCU ($1.50) line item gone, appended the price history row, and refreshed the repository numbers (147.4k stars, 24.7k forks, 16,908 commits).
See also #
- LlamaIndex - the retrieval-first counterpart, pivoting to documents
- Claude Code - the harness incumbent that dcode enters against
- Agentic Coding Tools Landscape - where framework vendors now collide with harness vendors
- The Importance of Context When Interacting with LLMs - the underlying reason retrieval and context beat raw model choice
References #
https://github.com/langchain-ai/langchain - repository scale (147.4k stars), MIT license, platform positioning, as of 2026-10-02
https://docs.langchain.com/oss/deepagents/code/overview.md - dcode, the terminal coding agent built on Deep Agents
https://www.langchain.com/pricing - LangSmith tiers and LSU metering, as of 2026-10-02
https://news.ycombinator.com/item?id=40739982 - the Octomind critique thread with the CEO’s response acknowledging over-abstraction
https://python.langchain.com/api_reference/text_splitters/text_splitters/code_splitter.html - current text splitters catalog, showing separator-based code splitting only