AivexaNewsSearch
AI news for builders and product teamsChecked every hour

How Workday Built an Enterprise-Scale AI Agent Platform with LangGraph and LangSmith

Collected Nov 10, 2026

Workday built an internal developer platform for creating, deploying and running AI agents on its own infrastructure, and the company credits LangGraph for orchestration and LangSmith for tracing, observability and evaluation. Alex Gatto, a Senior Machine Learning Engineer on the effort, described the results in a LangChain customer story. The push began in 2025, when Workday leadership identified an agentic market opportunity and set a goal of having agents in production by the end of 2026. Workday, which describes itself as an agentic platform for HR, finance and IT serving more than half the Fortune 500, needed tooling that could support both quick prototyping and production deployment under strict security and data governance rules.

The scope was large for a small team. The platform had to include a durable, scalable agent server; data access controls tied to Workday's audit policies; a shared evaluation platform; and a software development kit that abstracted all of it. The Agent Platform team serves two groups: internal Workday engineers who build agents, and the end customers who use those agents. The platform is meant to absorb authentication, security, compliance and infrastructure concerns so agent teams can concentrate on what an agent actually does.

The first major product to ship was Workday's Agent System of Record, a governance layer that controls which APIs an agent is permitted to reach. The Agent Platform team built the backend server and evaluation infrastructure beneath it. The team also assembled a Python SDK that Gatto calls "Workday LangChain": it wraps LangChain primitives and routes requests through Workday's internal services, including an LLM gateway and a RAG retrieval service. The pattern, as Gatto describes it, was to take each AI concept LangChain released and build an equivalent Python SDK version tailored to Workday.

LangGraph was chosen for orchestration because its graph-based model gives internal developers flexibility at the design layer while the platform supplies standardized infrastructure underneath. The comparison is instructive for anyone weighing agent frameworks: platforms that model agents as sequences of conversations or as personas push teams toward a particular shape, whereas a graph model lets engineers build anything from a simple ReAct loop to a complex deterministic workflow. In practice, a ReAct loop is an agent that alternates reasoning steps with tool calls until it reaches an answer; a deterministic workflow encodes fixed steps and branches. Graph abstractions accommodate both, which matters when a single internal platform must serve many teams with different needs.

Protocol support came essentially for free. When senior leadership asked about Workday's strategy for the A2A and MCP protocols, the team could say both were already in production, because the agent server implementations they ran were built with A2A and MCP support. Updates developers made to their graphs were reflected in the server automatically. Gatto's summary was that the backend was not the problem and LangChain handled it.

The headline number is time-to-prototype. Building a prototype agent previously took four to five months, with no guarantee it would ever reach production. On the new platform, a team can have a prototype running in a staging environment in under a day, with authentication and security requirements already baked in. That reframes the remaining work for an agent team as a design question: what the agent does and what its graph looks like. This is the most consequential claim in the story for other enterprises, because it suggests the binding constraint on internal agent development was infrastructure setup and compliance wiring rather than model capability.

The organization scaled alongside the platform. The team started with four people and now numbers roughly 30, with more investment planned. Gatto said the success of the LangChain partnership and what the team accomplished led Workday to more than triple its headcount. Gatto also framed the platform as more than a deployment layer. Over the years Workday engineers had built dozens of internal services, including LLM gateways, RAG retrievers, recommendation engines and prompt services, and most sat isolated, known only to their immediate teams. Through its Python SDKs and LangChain-backed primitives, the Agent Platform became a connective layer that surfaced those services, giving Workday a first unified interface for agent building that taps the depth of its internal infrastructure.

A few readings are worth drawing out for developers and product teams. First, the architecture separates concerns cleanly: orchestration is a developer-facing choice, while governance, identity and evaluation are platform responsibilities. That split is what makes a sub-day prototype possible without bypassing compliance, since security requirements are inherited rather than rebuilt. Second, the decision to wrap an open-source framework in an internal SDK is a common enterprise pattern, and it carries a trade-off: teams gain consistency and internal integrations, but they take on the maintenance of tracking upstream framework changes. Third, the A2A and MCP point illustrates a practical benefit of adopting protocol-aware infrastructure early, since answering leadership questions about interoperability becomes a matter of configuration rather than a roadmap item. MCP is an open standard for connecting agents to tools and data sources, and A2A is an open standard for agent-to-agent communication; support for both in the server layer means teams do not each reinvent it.

The prototype speedup also deserves a caveat. A staging prototype in under a day is not the same as a production agent, and the story does not quantify how long the remaining path to production takes. What it does establish is that Workday removed the setup phase that previously consumed months.

Why it matters: Enterprises building internal agent platforms now have a concrete reference for how much time sits in infrastructure and compliance plumbing rather than agent logic. Teams that adopt protocol-ready orchestration and a shared evaluation layer may find, as Workday reports, that prototyping collapses from months to a day, though the production path still needs its own scrutiny. The staffing shift, from four to roughly 30, also suggests the platform itself becomes a durable function rather than a one-off project, which is an inference from the reported growth rather than a stated plan.

Read at LangChain

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

See how Workday used LangGraph and LangSmith to build an internal AI agent platform that cut prototype time from months to less than a day and compliance needs.