LangGraph consulting that takes agents into production.

Soba Labs designs the graph and its state, builds the nodes and tools, adds a Postgres checkpointer, interrupts for approval, retry policies and streaming, deploys it and traces every run in LangSmith. We work in your repository with your team, and you own the code.

LangChain's create_agent returns a compiled LangGraph graph for the standard loop of model call, tool call, repeat, with middleware for approval, summarization and prompt changes. Deep Agents adds a file system for context, subagents and long-term memory for long, open-ended tasks. Write the LangGraph graph yourself when the topology is more than loop until done: classify before routing, fan out in parallel, run deterministic steps between model calls, or gate something that is not a tool call. All three run on the same runtime, so an agent from create_agent can be one node in a larger graph. See the LangGraph glossary entry for the basics.

With a Postgres checkpointer and a stable thread id per conversation or job, so every step is saved and a run resumes after a restart. We set the durability mode per graph: synchronous writes where a lost step would cost a duplicate side effect, background writes where throughput matters more, and a job that prunes old checkpoints so the table does not grow without limit. The Shen.ai case study shows session checkpointing in Postgres in a patient-facing assistant.

Tool calls get LangChain's human-in-the-loop middleware, which pauses before the tool runs. Anything else, a stage boundary or a threshold breach, gets an interrupt() call in its own node, resumed with a Command on the same thread. Because the node restarts from the top on resume, we keep writes before the interrupt idempotent or move them to the node after it. The reasoning is in LangGraph human-in-the-loop: interrupt vs middleware.

Usually not at first. A supervisor earns its place when there are several distinct domains, each with its own tools, and the subagents never talk to the user. Handoffs fit when agents take turns in one conversation. A router fits when the input categories are clear. Every extra agent is another prompt to maintain and another evaluation set to keep green, so we add one only when a single prompt can no longer hold its tools and instructions.

Either inside your own service around the compiled graph, or on LangSmith Deployment's Agent Server, which packages the graph with a persistence database and a task queue and exposes threads, runs and cron jobs over an API. It runs as managed cloud, hybrid with the agents in your VPC, fully self-hosted, or as standalone containers. We choose with your team based on where the data has to stay.

Every run is traced in LangSmith, failures from live traffic become examples in an evaluation dataset, and each prompt, model or graph change is scored against that dataset before it ships. The same loop lets us test an automatic prompt optimizer against a fixed task set before trusting its output, as in our Agent Lightning experiment.

Like every Soba Labs engagement: a fixed price on a defined outcome, agreed before we start. You own the graph, the checkpoints, the evaluators and the code.