By Keshavan Thiagaraja
Most enterprises have now moved past the question of whether to adopt AI agents. The harder question has arrived: how do you let autonomous software act on your behalf without surrendering control? The answer is not a bigger model. It is better architecture. Few stories illustrate this shift more clearly than the evolution from LangChain to LangGraph, and the lessons apply directly to any organisation building multi-agent systems today.
Why this matters now
Agentic AI has crossed from experimentation into production. In October 2025, LangChain raised 125 million US dollars at a 1.25 billion US dollar valuation in a Series B round led by IVP, with participation from Sequoia, Benchmark, CapitalG and the venture arms of ServiceNow, Workday, Cisco, Datadog and Databricks. The company reports around 90 million combined monthly downloads across LangChain and LangGraph, and states that over a third of the Fortune 500 use its services. Firms including Klarna, Cloudflare, Workday, Cisco and LinkedIn are reported by the company to run LangGraph in production.
Investors and enterprise buyers are not paying for chatbots. They are paying for the layer that makes agents reliable, observable and governable. That is the layer boards should be asking about.
What has happened
LangGraph was created by LangChain. It is not a separate company. The story begins with Harrison Chase, who studied statistics and computer science at Harvard, worked on entity linking at Kensho and then led machine learning testing work at Robust Intelligence. In October 2022 he released LangChain as an open source Python project that solved the unglamorous plumbing of building applications on large language models: connecting models to tools, data sources and workflows.
In early 2023 Chase co-founded the LangChain company with Ankush Gola, a Princeton-trained engineer who had spent four years at Facebook and was a former colleague at Robust Intelligence. Chase drove the product vision of helping developers build reliable AI agents that can use tools, retain state, make decisions and run complex workflows. Gola partnered in building the company around the open source project, including its commercial and operational growth.
The sequence that followed is instructive for every technology leader. LangChain began as a framework for building LLM applications. As adoption grew, developers hit a wall: simple prompt chains could not support long-running, reliable and controllable agent workflows. In January 2024 the company answered with LangGraph, a graph-based runtime for stateful AI agents that supports human approval steps, retries, persistence and complex orchestration. In October 2025 both LangChain and LangGraph reached 1.0 stability milestones, with the LangChain framework itself rebuilt on top of the LangGraph runtime. The framework that started the ecosystem now runs on the orchestration engine it created.
The technology and architecture implications
The move from chains to graphs is more than a naming change. A chain is a straight line: input flows to output through a fixed sequence. A graph is a map of decisions: an agent can branch, loop, retry, delegate to specialist agents and pause for human input. For enterprise architects, three properties matter most.
First, state becomes a first-class citizen. LangGraph persists the state of every agent workflow through checkpointing, so a process can pause for hours awaiting approval and resume exactly where it stopped. Second, human-in-the-loop control is built into the runtime rather than bolted on. An approval step is a node in the graph, not an afterthought in the user interface. Third, orchestration patterns become explicit. A supervisor agent can coordinate specialist agents in a hierarchy, with clear contracts about who may do what.
In practical terms, agent orchestration is emerging as a new middleware tier in the enterprise stack, sitting between foundation models and business systems, in the way application servers once sat between databases and user interfaces. Observability platforms such as LangSmith complete the picture by making agent behaviour traceable and testable.
The business and competitive implications
Reliability, not raw intelligence, is the competitive moat in enterprise AI. Consider an insurance claims operation as an illustration. A claims platform can use a supervising orchestrator agent to control specialist agents for document intake, policy verification, damage assessment and fraud screening. The system can assemble the evidence overnight, validate the policy terms, estimate the settlement and flag anomalies for investigation. What it cannot do is release a payment. That action sits behind an approval gate, and only an authorised claims officer can trigger it.
This is precisely the pattern LangGraph was designed to support, and it is why the framework is especially relevant in regulated industries such as insurance, banking and healthcare. The commercial logic is simple: the value of automation compounds only when the cost of a wrong action is contained. Organisations that master this pattern can automate aggressively where errors are cheap and gate decisively where errors are expensive. Competitors that treat agents as unsupervised automation will either move too slowly out of fear or move too fast and pay for it publicly.
The risks and governance considerations
Four risks deserve board-level attention. Autonomy without gates: an agent that can act on external systems without approval checkpoints is an operational and reputational liability. Auditability: persistent state and full traces are assets for compliance, but only if retention, access and privacy controls are designed in from the start. Concentration risk: LangGraph is open source under an MIT licence, which limits lock-in at the framework level, but commercial platform services around deployment and observability still create dependency that should be assessed like any strategic vendor relationship. Regulatory expectation: supervisory regimes worldwide increasingly expect demonstrable human oversight of consequential automated decisions, and approval gates are among the clearest ways to evidence it.
What CEOs should do
1. Ask where agents in your organisation can act, not just what they can answer. The risk profile of an agent that reads is entirely different from one that transacts.
2. Demand a written approval-gate policy: which decisions agents may take alone, which require human sign-off, and who is accountable for each gate.
3. Fund reliability engineering, not demonstrations. Budget for observability, evaluation and state management as core costs of agentic AI, not optional extras.
4. Treat orchestration frameworks as strategic infrastructure decisions with multi-year consequences, and involve the board risk committee early.
What CTOs should do
1. Standardise on an orchestration layer before teams proliferate bespoke agent stacks. Evaluate LangGraph alongside alternatives against your control, persistence and audit requirements.
2. Design human-in-the-loop as architecture. Approval steps should be durable nodes in the workflow graph with full state checkpointing, not front-end conventions.
3. Instrument everything. Every agent decision should produce a trace that compliance and engineering can replay.
4. Start with asymmetric workflows such as procurement, claims or vendor management, where research and recommendation can be automated safely while commitment actions stay gated.
5. Build exit optionality. Keep agent logic portable and document the boundary between open source framework and commercial platform services.
The strategic conclusion
The LangChain to LangGraph journey mirrors the journey every enterprise must now make: from impressive prototypes to governed production systems. The winners in the multi-agent era will not be the organisations with the most agents. They will be the organisations that know exactly what their agents are allowed to do, can prove it, and can scale that confidence across the business.
Engagement question
Where would you draw the line in your organisation between what an AI agent may decide on its own and what a human must approve? I would value your perspective in the comments.
Think mathematically. Engineer intelligently. Lead strategically. Build responsibly. Scale globally. Create real business value through artificial intelligence.
#ArtificialIntelligence #AgenticAI #MultiAgentSystems #LangChain #LangGraph #EnterpriseAI #AIArchitecture #TechnologyStrategy #ResponsibleAI #DigitalTransformation #CTO #CEO
Originally published on LinkedIn by Keshavan Thiagaraja. Read the original on LinkedIn.
