We design, build and harden agents on LangChain's open-source Deep Agents harness: planning, subagents, a virtual filesystem and durable runs on LangGraph, with LangSmith tracing and evaluations from the first week. Built by official LangChain Ambassadors and Experts, in your environment.
LangChain experts
Deep Agents bundles a task list, subagents, a virtual filesystem and context compression on top of LangChain's agent loop. We start by checking whether the job needs that: many agents are better served by create_agent or a custom LangGraph graph, and we say so before anything is built.
Research, coding and verification run as isolated subagents with their own context, and intermediate results live in the virtual filesystem instead of the conversation, so a run of hundreds of steps keeps its instructions in view.
Agent-written code runs in an isolated sandbox backend, never on the host, and sensitive tools pause for a person to approve, edit or reject through the harness's interrupt configuration, with the run resuming from its checkpoint afterwards.
Every subagent call, file write and tool call is traced in LangSmith, and a dataset of real tasks scores the whole trajectory, so a change to the instructions, skills or model ships only after it passes the same checks as the last one.
Managed Deep Agents takes a code-first project (instructions, skills, tools, subagents) and runs it on LangSmith's Agent Server with durable threads, sandboxes, memory, approvals and automatic tracing, deployed from the mda CLI. As of October 2026 it is in public beta on LangSmith Cloud, in the US region only, and the docs offer no self-hosted or hybrid option for it. We can build on it when that fits, and we tell you when it does not.
A team that has to keep data and compute in Europe today runs the same harness on its own infrastructure: the open-source deepagents package, a persistent LangGraph checkpointer and a sandbox provider you choose, with LangSmith tracing pointed at the EU region or a self-hosted instance. The agent definition carries over unchanged if you move to the managed runtime later.
Soba Labs is based in Poland. We work in your cloud account and your repositories, and you own the agent definition, the evals and the code whichever runtime it ends up on.
Deep Agents is LangChain's open-source agent harness, the deepagents package, built on top of create_agent and the LangGraph runtime. It bundles a planning task list, subagents, a virtual filesystem, automatic context compression and skills, so an agent can carry a long, multi-step task such as research or coding without losing track of what it has done. The glossary entry explains the four layers.
When the task is short and well defined. A support reply, a classification step, a structured extraction or a fixed pipeline does not need a task list, subagents or a filesystem, and the extra prompt and tool surface costs tokens and adds failure modes. LangChain's own guidance is to use create_agent for a lighter harness and a custom LangGraph graph when the agent loop itself is the wrong shape. We recommend Deep Agents only when the work is long, varied and benefits from delegation.
Durable execution first: a persistent LangGraph checkpointer so a failed or paused run resumes from its last step instead of starting over. Then a sandbox backend for any code the agent writes, interrupt configuration on every tool with a side effect so a person approves it, LangSmith tracing on every run, and an evaluation dataset of real tasks that gates each change to the instructions, skills or model. Memory and skills are reviewed as code, in your repository.
Managed Deep Agents is LangChain's hosted runtime for Deep Agents: you keep the agent definition in your repository and deploy it with the mda CLI, and LangSmith runs it on its Agent Server with durable threads, sandboxes, memory, human approvals and automatic tracing. As of October 2026 LangChain's docs state it is in public beta on LangSmith Cloud in the US region only, with no self-hosted or hybrid option. A European team that must keep data in the EU today runs the open-source harness on its own infrastructure and can move the same definition to the managed runtime later.
Like every Soba Labs engagement: a fixed price on a defined outcome, agreed before we start. You own the agent definition, the evals and the code.
The reasoning core of a multi-agent localisation pipeline: a structured client rule base, the agents that reason over it, and producer approval at key decision points.
End to endfresh jobs passed every step, from brief to packaged multi-market, multi-format deliverables
Read case study
A production sales agent that qualifies inbound leads, answers product questions from a live knowledge base, and writes structured outcomes back to the CRM.
No duplicatesone CRM write per conversation turn, held by idempotency keys and a timeout regression test
Read case study
A patient-facing assistant that explains health scan results from an approved medical knowledge base. Low-confidence and clinical questions route to a clinician.
15%of conversations routed to a clinician; the rest answered from approved material
Read case study1/1
We sell a fixed price on a defined outcome, delivered in days. Every system is custom to your process and your data, and you own it and its source. No product to license, no platform to rent, no lock-in.
Book a consultation