Rippling Builds Multi-Agent AI Layer for Workforce Management Using LangChain Deep Agents

Rippling is a workforce management platform that manages everything from onboarding and benefits to device provisioning and spend management. Its data model spans HR, IT, payroll, finance, and global operations, encompassing thousands of tables, hundreds of thousands of fields, and concepts that share names across domains. Building an AI layer that reasons across all of it required a new architecture. Rippling AI, now in production across millions of users globally, runs on LangChain Deep Agents and LangSmith. The team shipped it in roughly six months, using a supervisor agent that coordinates five to seven specialized subagents: read agents that query structured data across Rippling's product areas, RAG agents that retrieve from help center docs and HR policy documents, and action agents that execute write operations such as uploading bonuses, normalizing job titles, or triggering new hire onboarding. The core technical challenge was context engineering at scale. The team developed three patterns: dynamic skill injection via Deep Agents middleware to reduce context bloat by 100 to 500x, sandboxed code execution for write operations to separate LLM reasoning from deterministic data normalization, and a REPL variable store to prevent hallucination of long alphanumeric IDs. LangSmith provides the shared observability layer, running a layered eval system with offline mocks, 300 to 400 post-merge integration queries, deploy-blocking evals against real systems, and continuous production monitoring multiple times daily. A semi-automated self-healing loop pulls failing traces, has an agent propose fixes, and re-runs evals until regressions close. Rippling's product owner Laks Srini noted that LangSmith makes it possible to pull and analyze all conversations at scale, with automated analysis running on top of it.