Lyft Builds Self-Service AI Agent Platform with LangGraph and LangSmith

Lyft's existing AI agent development approach was overly dependent on machine learning engineers (MLEs), creating a bottleneck where development cycles took several months. To solve this, Lyft built a self-service platform that allows non-technical experts such as operations teams and product managers to directly define and refine agents through natural language.

The system architecture was designed based on LangGraph's Router Multi-Agent pattern.

• Meta Agent: Maintains state, classifies requests, and passes tasks to the appropriate subagent via Command(goto=...).
• Subagent: Consists of specialized graphs that handle Rider- or Driver-specific intents.
• Handoff: If a more specialized agent is needed mid-conversation, control can be returned to the meta agent for rerouting.

To maintain agent reliability and quality, Lyft introduced a monitoring system leveraging LangSmith. It performs automated validation through an LLM-as-a-judge evaluation system, and precisely tracks agent behavior through tracing and dashboards.

Through these changes, Lyft achieved a reduction in agent development time from about 6 months to just a few weeks, and established an environment where domain experts can solve problems directly without technical intermediaries.