LangChain's case study says Rippling AI, now used by more than one million people globally, runs on LangChain Deep Agents and LangSmith and was shipped in roughly six months. The architecture is a supervisor agent coordinating 5 to 7 specialized subagents. Dynamic skill injection reduces context size by 100 to 500x, and the team runs 300 to 400 post-merge integration queries against a full Rippling sandbox before deployment.
Laks Srini, Product Owner, Rippling AI, said: "As we embedded AI into individual products, it became clear that siloed, vertical-specific models couldn't scale."
Rippling's data model spans HR, IT, payroll, finance and global operations: thousands of tables, hundreds of thousands of fields, and concepts that share names across domains. A question such as "What's my balance?" could pertain to a health savings account, a credit card, a contractor payment account or a time-off policy, and passing schema chunks to an LLM does not work with that many overlapping entities.
The team built with LangChain from the start, using composable agent primitives and a shared observability layer in LangSmith, and adopted Deep Agents for Rippling AI's core reasoning loop when it launched. Customers use a chat interface inside the Rippling portal and mobile app. Beneath the supervisor agent sit three types of Deep Agents. Read agents query structured data across HR, payroll, IT and finance and connected platforms such as Salesforce, Carta and GitHub. RAG agents retrieve from help center docs, company handbooks and HR policy documents. Action agents execute write operations, such as uploading bonuses, normalizing job titles and leveling structures, or triggering new hires pre-populated from prior employee profiles.
Context engineering was the core technical challenge, and the team developed three patterns. With dynamic skill injection, a search step uses Rippling's semantic layer to identify the relevant domain, then injects a skill scoped to it, such as payroll, devices, ATS or spend. Action agents use sandboxed code execution to normalize inputs, such as a CSV from a client, separating LLM reasoning from deterministic formatting. A REPL keeps a runtime variable store between agent steps, so agents refer to named variables because LLMs hallucinate when reciting long alphanumeric IDs.
LangSmith runs a layered eval system: offline evals on every commit, post-merge integration evals, about 10 deploy-blocking critical scenarios, and continuous evals on production data multiple times daily. A semi-automated self-healing loop pulls failing production traces, has an agent propose fixes and re-run evals until regressions close, and a human reviews and merges the resulting PRs.
Details