Agents that chain several tools
An agent that looks up a file, queries the ERP, drafts a reply and creates a ticket, with different rules at each step.
AI integration · Agent frameworks
LangChain and LangGraph Development: AI Agents and Orchestration in Production
Etixio builds AI agents and LLM pipelines with LangChain and LangGraph: multi-step orchestration, tool calls, memory, human approval and recovery after errors. We connect them to your data and tools, then take them to production with observability, evaluations and cost tracking. And we tell you when a framework is not needed.
Understanding the technology
Two open-source frameworks, two levels of abstraction.
LangChain provides building blocks for applications based on language models: a common interface to model providers, tool definitions, structured outputs, and integrations with vector databases and data sources. Since version 1.0, it offers an agent creation function that covers the common case: a model that calls tools in a loop until it produces its answer.
LangGraph is the low-level orchestration layer these agents run on. A process is described as a graph of steps sharing a state: each node performs an operation, transitions can be conditional and the state is saved at every step. This makes it possible to resume an interrupted process, wait for human approval or have several agents work together. Both frameworks are available in Python and JavaScript/TypeScript.
Business use cases
When an orchestration framework adds value.
Processes with human approval
Handling a request that stops before a sensitive action, waits for a manager’s approval, then resumes where it left off.
Multi-pass document processing
Classification, extraction, consistency checks and then summarization, with each step able to use a different model.
Multi-agent systems
A coordinator agent that splits work among specialized agents, when the task genuinely calls for it.
Our scope of work
What we put in place.
Orchestration and state
Design of the processing graph, shared states and transition conditions, with deterministic steps wherever the model is not needed.
Tools and connections
Tools connected to your APIs, your RAG document search or MCP servers, with parameter validation and permission checks on the application side.
Memory
Conversation memory persisted in a database (for example PostgreSQL) and, when needed, long-term memory per user, with retention and deletion rules.
Human approval and recovery
Breakpoints before sensitive actions, resumption after interruption, and handling of errors, timeouts and retries.
Observability
Traces of every run (steps, model calls, tools, tokens consumed, duration) in LangSmith or an equivalent such as Langfuse, depending on your data hosting constraints.
Evaluations
Sets of representative scenarios rerun whenever the model, instructions or graph change. See our approach to LLMOps and AI evaluation.
The choices that matter
When not to use a framework.
A framework adds dependencies, layers of abstraction and version upgrades to keep up with. For a single model call, a structured extraction or an assistant that uses only a few tools, the provider’s SDK (Claude, OpenAI, Gemini or Mistral AI) plus some application code is often easier to understand, test and maintain.
LangGraph becomes worthwhile when the process involves several steps with state to preserve, interruptions, resumptions or multiple agents. We make this call during scoping, based on the real complexity of the process and the skills of the team that will maintain the service. An agent already built with a framework can also be simplified if the abstraction makes it harder to operate.
Security, monitoring and costs
Taking a LangGraph agent to production.
Security
The model does not decide on its own what it is allowed to do. Permissions are checked by the tools, write actions are kept separate and content read by the agent is treated as untrusted data.
Monitoring
Alerts on errors, loops and abnormal durations, plus dashboards tracking quality and user feedback.
Costs
Caps on the number of steps and on usage per run, a lighter model for simple steps, caching, and cost tracking per request type.
Deployment
A containerized service in your cloud or on your infrastructure, with its environments, state backups and operations documentation.
In the field
Agents connected to real data.
Our AI case studies show the same principle, whatever tooling is chosen. The medical AI agent explores the patient record and returns a chronological summary to anesthesia teams; Bobby connects a conversational assistant to our ERP data. In both cases, the value lies as much in the data connection, the controls and the interface as in the model.
From work to deliverables
What we deliver.
- Agent architecture and the rationale for the tooling choice (framework or not).
- Code for the agent, its tools and its connections, with tests.
- Observability, evaluation set and measured results.
- Deployment, monitoring, cost tracking and documentation.
A LangChain prototype can be taken over and industrialized; see our AI POC to production offering our AI agents and chatbots and our AI features for your product. The work is delivered as a fixed-price project or with a dedicated team.
Frequently asked questions
LangChain and LangGraph: your questions.
What is the difference between LangChain and LangGraph?
LangChain provides the building blocks and a fast way to create a standard agent. LangGraph is the underlying orchestration engine, which lets you describe a multi-step process precisely, with saved state, branches, interruptions and resumptions. Teams often start with LangChain and move down to LangGraph when the process requires it.
Do you need LangChain to build an AI agent?
No. A simple agent can be built directly with the model provider’s SDK. A framework makes sense when orchestration becomes complex or when you want to switch providers easily. We compare both options during scoping.
What is LangSmith?
LangSmith is LangChain’s platform for tracing, evaluating and deploying agents. It is not mandatory; tools such as Langfuse can collect traces from a LangGraph application, including self-hosted if your data requires it.
Does LangGraph work with Claude, OpenAI or Mistral?
Yes. LangChain offers integrations for the major model providers, including Anthropic, OpenAI, Google and Mistral AI, as well as for hosted models. The model is chosen based on your examples, and several models can be used at different steps.
Can you take over an existing LangChain agent?
Yes. We review the code, the versions in use and the results achieved, update dependencies if needed, add evaluations, observability and security controls, then decide with you whether to keep the framework or simplify.
Let’s build a maintainable AI agent.
Tell us about the process to automate, the tools involved and your technical environment. Together we will define the initial scope to explore.