AI BriefingRippling Builds Production AI in 6 Months with Deep Agents and LangSmith
Key point
Rippling used Deep Agents and LangSmith to build production AI handling complex data models in just 6 months.
Details
Rippling has a massive data model spanning thousands of tables across HR, IT, payroll, and finance, with overlapping and duplicated concepts. Simply feeding the schema to an LLM wasn't enough to resolve ambiguity across domains and answer complex questions.
To solve this, Rippling adopted a multi-agent architecture built on LangChain Deep Agents. A Supervisor agent sits at the center, orchestrating 5–7 specialized sub-agents, enabling the team to bring AI to production in about 6 months.
The system consists of three types of specialized agents:
- Read agents: Query structured data across HR, payroll, IT, etc., as well as external platforms (Salesforce, GitHub, etc.)
- RAG agents: Retrieve unstructured data such as help docs and internal handbooks
- Action agents: Perform write operations within the system, such as uploading bonuses or adjusting job titles
LangSmith handles production debugging and evaluation. It leverages tracing, hierarchical evaluation, and a semi-automated self-healing loop to continuously improve system quality, preventing model performance degradation and managing quality.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.