Remote Deploys LangChain AI Agents to Automate Thousands of Customer Migrations

Remote, the global employment platform, has deployed an AI-powered Code Execution Agent built on LangChain and LangGraph to automate customer data migrations that previously required manual scripting for each new client.

The system addresses a fundamental constraint in enterprise AI: context windows. Even GPT-5’s 400,000-token limit falls far short of what’s needed to process large payroll spreadsheets, which can exceed 50MB. Feeding that data directly into an LLM produces hallucinations—a non-starter for a platform handling sensitive employment and compliance data across multiple jurisdictions.

Separating Reasoning From Execution

Remote’s solution splits the workload. The LLM handles reasoning and task planning through LangChain’s tool-calling interface. Actual data manipulation happens in a sandboxed Python environment running WebAssembly, where Pandas handles the heavy lifting.

“Large intermediate results never pass back to the model, keeping the context small,” wrote José Mussa, Staff Software Engineer at Remote, in a technical breakdown of the system.

The workflow runs like this: customers upload raw data exports, the agent maps out transformation steps, sandboxed code executes those transformations, and the agent iterates until output matches Remote’s employee schema. The proof of concept processed a 5,000-row Excel file, mapping entries to schema and answering queries by running code rather than generating text.

Why Graph-Based Orchestration

LangGraph’s node-and-edge architecture proved critical for managing multi-step migrations. Each workflow stage—ingestion, mapping, execution, validation—becomes a discrete node with explicit transitions for success, failure, or retry. This makes agent state transparent and recoverable, borrowing patterns from distributed systems engineering.

The framework choice aligns with broader industry adoption. According to LangChain’s 2024 state of AI report, 43% of LangSmith organizations now run LangGraph traces. The Python and JavaScript versions have crossed 130 million combined downloads.

Production Results

Remote’s onboarding teams no longer write custom scripts per customer. The agent transforms diverse formats into consistent JSON schemas in hours instead of days. Because transformation logic executes as auditable code rather than LLM-generated text, the output is repeatable—essential for regulatory compliance across employment jurisdictions.

The team extracted several operational principles from the build. LLMs function best as planners, not data processors. Graph-based workflow structure simplifies debugging. Context tokens should be treated as scarce resources, with large intermediate results staying in the execution environment.

Expanding the Pattern

Remote is now abstracting similar patterns into reusable agents across the company. A recent prototype combines document parsing with agentic workflows for OCR-to-JSON conversion, reportedly outperforming basic OCR by a significant margin.

The company plans to contribute generic improvements back to LangChain’s open-source ecosystem. For enterprise AI teams wrestling with similar data transformation challenges, Remote’s architecture offers a template: let models think, let code execute, and keep the two cleanly separated.

Image source: Shutterstock