Lyft reduced AI-agent development from months to weeks with LangGraph and LangSmith

Lyft created a self-service platform for developing AI agents that help with customer support. Using LangGraph and LangSmith tools, the company reduced agent development time from months to weeks, making AI innovation accessible to all engineers, not just a narrow group of machine learning specialists.

The

Problem: Development Was a Bottleneck Previously, creating a single AI agent for Lyft required several months of work from a specialized team. Deep knowledge of machine learning, integration with Lyft systems, monitoring setup, and continuous improvement were necessary. This made the process expensive and slow — a new idea from hypothesis to production took a quarter or two.

The company serves millions of rides a day, and customer support is a huge volume of work. Every incorrect chatbot answer can lead to an unsatisfied customer or additional load on the support team. The team realized that if they could empower developers to independently create and test agents, the gains in speed and quality would be significant.

But for that, they needed the right tools to make development intuitive.

Solution:

LangGraph for Logic, LangSmith for Visibility Lyft chose LangGraph to orchestrate agent logic and LangSmith for monitoring, debugging, and testing. LangGraph allows you to describe an agent's workflow as a graph: nodes are individual steps (searching information in a database, generating a response, checking for relevance, escalating to a specialist), edges are conditional transitions between them. This is more intuitive to an engineer than a set of callbacks and asynchronous code.

LangSmith adds process visibility: logging all steps, quality metrics, A/B testing of prompts, quick error tracking. The team can immediately see where the agent got confused, which prompt didn't work, and fix it on the fly without redevelopment. This is critical because when working with LLMs, errors often hide in the details of wording.

Platform's main features: Fast iteration on prompts and logic without service reload Graph approach simplifies adding new steps to the workflow Built-in monitoring saves weeks of debugging and log analysis Ready-made templates for common tasks (search, ranking, routing) ## Practical Result: Weeks Instead of Months After implementing the platform, time from idea to agent in production dropped from 8–12 weeks to 1–2 weeks. Now a product manager can suggest an idea, an engineer assembles a workflow in LangGraph in a day or two, tests it on real data through LangSmith, looks at metrics, and the agent is ready to launch. The work shifted from deep ML to logic design, which is more accessible to most of the team.

Lyft has already launched several agents on the platform: from simple (automated FAQ answers) to complex (triage systems that decide which specialist to escalate to, assess priority, find relevant documentation). Each operates with lower response latency and better accuracy than old rule-based systems.

What

This Means AI platforms are becoming tools not just for ML engineers, but for entire engineering teams. When tools are simple to learn and results are visible immediately, the speed of AI adoption in business grows several-fold. For companies at Lyft's scale with millions of inquiries a day, this means rapid experimentation with new AI solutions, measuring real impact, and scaling those that work.