Key findings
- 26 of 29 sources that discuss explicit control flow are positive, including 12 of 13 team and leadership sources.
- Checkpointing earns praise in 11 of 19 sources. The 8 critical accounts describe lost state, checkpointer bugs, tool calls repeated on resume, memory growth under async durability, or persistence that took extra work.
- Complexity and overhead lead the concerns with 11 mentions. Breaking changes and regressions appear in 7 sources, most from before 1.0.
- Of the 25 sources that describe a framework choice, 16 chose or kept LangGraph and 9 replaced it, 6 of them with their own code.
- Company-wide deployments include LinkedIn's framework across 20+ teams and BlackRock's Aladdin agents with tools from more than 50 teams. Toyota reports more than 50 production agents.
Teams keep LangGraph for explicit control flow and checkpointing. Most of the teams that replace it move to their own code after using a graph for work that did not branch. GILD read English-language public accounts published from March 2025 through October 2026, including talks by LinkedIn, BlackRock, Toyota, Airbnb and LATAM Airlines and engineering write-ups from Qodo and Wayfair.
What teams keep it for
Teams that stay name capabilities they did not want to build or maintain. Airbnb tried other frameworks and chose LangGraph for data access, access controls, stop and replay, and per-run inspection, and because a homegrown framework would leave its own engineers carrying the pager. Bridgewater moved its chat agent to LangGraph for persistent state and built-in cancel and continue, which it had managed itself "with a much worse effect". The Pizza Bot team threw away a homegrown checkpointing layer that kept producing edge-case bugs.
In large companies, one LangGraph setup serves many teams. LinkedIn builds its GenAI apps on LangChain and LangGraph behind its own adapters and says prebuilt agents cut new apps from weeks to days. BlackRock's Aladdin platform takes tools and agents from more than 50 engineering teams through a plugin registry, and every one of Toyota's more than 50 production agents is a config file over generated LangGraph graphs, which LangChain says cut delivery from six months to four days.
How production teams use the graph
Teams use deterministic rules to constrain model decisions. Airbnb runs a rules layer before any LLM call and lets agents abstain when a person must decide. Its first version abstained on about 25% of cases it should have approved. A Salesforce engineer's Switchplane project calls the LLM only for judgment. Wayfair's ticket-flagging graph identifies the comment author in code before any model call, and splitting one prompt into three author-specific prompts raised accuracy from 70% to 96%.
Production agents also wait for people. Airbnb pauses trust cases for days with interrupts and stores case data in its database outside graph state. Another team adds a human checkpoint after every critical step.
Where it adds work
Overhead leads complaints. Developers cite boilerplate for simple loops and a steep learning curve. Rippling replaced a sub-agent design with one flat agent after handoffs and shared context became hard to manage alongside interrupts. Apollo.io found a supervisor with hardwired subagents hard to extend. Other accounts describe runtime errors from loosely typed state and TypeScript lagging Python in 2025.
Breaking changes form the second theme, mostly before 1.0. Version 0.4.4 broke a basic import. A Postgres checkpointer release could not read saved state, and course notebooks failed within a week. LangGraph 1.0 shipped on October 22, 2025 with a commitment to no breaking changes until 2.0.
Why some teams replaced it
Six of the nine replacements moved to the team's own orchestration code. yeos.ai removed LangGraph after months because debugging across state transitions got harder as customers grew and most requests were sequential. One team chose Pydantic AI over what it saw as LangGraph's deployment lock-in, and a developer who used LangGraph in 8 of 10 projects moved most of them to simpler code.
Several who left still credit LangGraph for branching, stateful work and say their mistake was using it for pipelines that did not need it. Ask these questions before you choose.
- Does the workflow branch, wait for people, or need to resume after failure?
- Who will debug the graph in production, and with what tracing?
- Which LangGraph and checkpointer versions are pinned, and who owns upgrades?
- Is the orchestration layer behind your own interfaces, so it can be replaced?
Methodology
GILD opened 57 sources on October 12, 2026, 6 on October 25, 2026, and 4 on October 26, 2026. All were published between March 2025 and October 2026. They include Hacker News comments, GitHub issues on LangChain's LangGraph repository, company engineering blogs, and named developers' dev.to posts. Others are the InfoQ transcript of a QCon London 2025 talk and customer talks at LangChain's Interrupt conference. Every count in this piece counts those 67 sources, written by engineers who chose to post in public. On October 28, 2026 GILD added 13 more, from Tiendanube's and LinkedIn's engineering blogs, 8 Reddit threads, 2 GitHub issues and a Software Engineering Daily podcast; they appear on the Signals page and are not in this piece's counts. GILD left LangChain's own blog and customer stories out of the counts and checked product behavior against LangChain's documentation on the same dates.
Products in this piece
Sources
- Public source
- 21
- Official documentation
- 4
- Farshid Bahrami and colleagues (Wayfair Tech Blog): Flag Early, Fix Faster: Real-Time Intervention for Supplier Transfers with Wilma (Jan 15, 2026)
- Pedro Rodriguez (Airbnb): How Airbnb builds Trust for Guests and Hosts using AI Agents with LangGraph, Interrupt NYC 2026 (Sep 29, 2026)
- Brendan McManus, Michael Ran, and Santi Weight (Bridgewater): How Bridgewater Built an AI Analyst, Interrupt 26 (Jul 24, 2026)
- Hacker News comment on replacing homegrown checkpointing with LangGraph in Pizza Bot (Sep 16, 2026)
- David Tag (LinkedIn): How LinkedIn Built Their First AI Agent for Hiring with LangGraph, Interrupt 2025 (Jun 13, 2025)
- Karthik Ramgopal and Daniel Hewlett (LinkedIn): Lessons Learned from Building LinkedIn's First Agent, QCon London 2025, transcript on InfoQ (Dec 31, 2025)
- Brennan Rosales and Pedro Vicente Valdez (BlackRock): How BlackRock Built Production AI Agents with LangGraph, Interrupt 2025 (Jun 12, 2025)
- Ravi Chandu Ummadisetti and Kordel France (Toyota): Building ToyotaGPT, Interrupt 26 (Jul 15, 2026)
- Hacker News comment on Switchplane, LangGraph control flows that call the LLM only for judgment (May 4, 2026)
- Hacker News comment on a human checkpoint after every critical step (Jan 23, 2026)
- Senthil Velu Sundaram and Akash Ashok (Rippling): Rippling's Playbook, Interrupt 26 (Jul 13, 2026)
- Aniruddha Laud and Anshul Pahwa (Apollo.io): From Supervisor to Deep Agents, Interrupt NYC 2026 (Sep 29, 2026)
- Hacker News comment on runtime type errors from loosely typed LangGraph state (Jun 18, 2025)
- Hacker News comment on LangGraph's TypeScript version lagging Python (Nov 25, 2025)
- LangGraph issue #4706: 0.4.4 ModuleNotFoundError for langgraph.cache (May 15, 2025)
- LangGraph issue #5677: AsyncPostgresSaver 2.0.23 fails to read channel values (Jul 27, 2025)
- Nicola Richli (yeos.ai) on dev.to: Why We Removed LangGraph From Our AI Platform (Jul 27, 2026)
- Hacker News comment choosing Pydantic AI over LangGraph on deployment lock-in (Mar 24, 2025)
- DeadLocker on dev.to: Why I Stopped Using LangGraph (Apr 23, 2026)
- Hacker News comment on moving from LangGraph and the Agents SDK to an in-house framework (Oct 8, 2025)
- Sagi Medina (Qodo): Why we chose LangGraph to build our coding agent (Mar 21, 2025)
- LangChain blog: LangChain and LangGraph reach v1.0 (Oct 22, 2025)
- LangChain docs: Release policy (LangGraph LTS)
- LangGraph docs: Checkpointers and durability modes
- LangGraph docs: Interrupts