Workday’s Agent Platform Cuts Agent Prototyping from Months to Under a Day
- Cargo
- Machine Learning
- Shipped
- Shipper
- Joseph Rodriguez
- Readers
- 645
A four person team built an internal platform around LangGraph and LangSmith, reducing agent prototyping time while standardizing access to Workday services.

Workday says its internal agent platform reduced the time to get an agent prototype running in staging from four to five months to less than a day. That change matters because a prototype at a large enterprise needs a path to production: developers must work within existing authentication, security, audit and data retention requirements from the start.
The platform combines LangGraph for agent orchestration with LangSmith for tracing, observability and evaluations. Workday built the surrounding services and Python SDK to connect those tools with its internal infrastructure. The result gives agent developers room to define different workflows while a shared platform handles much of the operational setup.
A small team built the first version
Workday began the effort in 2025, when its leadership set a goal of getting agents into production by the end of 2026. The team had to support both quick experimentation and production deployment, while meeting the company’s demanding controls for internal services. Its initial scope included a scalable agent server, access controls tied to Workday’s audit policies, an evaluation platform for teams, and an SDK that hid much of that infrastructure behind a developer interface.
Four people took on that work. The platform serves two groups: Workday engineers building agents and the customers who use those agents. For internal developers, the aim is to keep attention on an agent’s behavior and intended function. The platform supplies shared capabilities for authentication, security, governance and infrastructure.
One of the first products was the Agent System of Record, a governance layer that determines which APIs an agent can call. The platform team built the backend server and evaluation infrastructure on which that system runs. API access rules are central to the design: an agent’s capabilities depend not just on its graph, but also on which Workday services it is permitted to reach.
An SDK connects existing services
Workday also developed a Python SDK that Alex Gatto, a senior machine learning engineer, calls “Workday LangChain.” It wraps LangChain primitives and routes them through internal services, including an LLM gateway, a retrieval service for retrieval augmented generation, and other tools. Gatto said the team aimed to provide Workday versions of the AI concepts that LangChain introduced, giving engineers a consistent way to use those capabilities inside the company.
This layer addresses a problem that can emerge as large organizations build AI infrastructure in separate teams. Workday had accumulated services for language models, retrieval, recommendations and prompts, but many were isolated from one another and known mainly to their immediate users. The platform’s SDK and shared primitives give agent builders a common way to reach those services. That makes the platform a connection point across existing infrastructure, as well as a place to deploy agents.
Graphs leave room for different agent designs
Workday chose LangGraph because its graph model lets developers define a broad range of agent behaviors without requiring the platform to prescribe one design. Engineers can build ReAct loops or deterministic workflows, then run them on standardized infrastructure. That separation gives product teams flexibility in how they implement an agent while keeping deployment and operations consistent underneath.
Gatto pointed to LangChain’s Agent Server and LangSmith as tools that include capabilities the team would otherwise have needed to build and maintain. One practical example was support for A2A and MCP protocols. When Workday leaders asked about the company’s approach to those protocols, Gatto said the team could report that support was already in production. The agent server implementations the team used included that support, so it was available across the servers it ran.
Updates to a developer’s graph also appeared in the server without requiring the platform team to build a separate backend for each change. “The backend isn’t the problem,” Gatto said, crediting LangChain with handling that infrastructure. This can shorten the path from an agent design to a working deployment, though the account does not specify the volume of agents served, the underlying compute requirements or the cost of operating the platform.
Prototype speed changes the production path
Before the platform, a prototype could take four to five months to build, and there was no guarantee it would be suitable for production. Workday now says a team can have a prototype running in a staging environment in under a day, with authentication and security requirements already included. Developers can spend more of their time deciding what the agent should do and how its graph should work.
That result is about prototyping speed, not a reported reduction in the time required to launch a finished customer product. Moving from a staging prototype to production still depends on the agent’s behavior, its permitted access to services and the organization’s deployment requirements. The platform’s value is that teams can begin that work within a shared environment instead of first assembling the underlying infrastructure themselves.
The team has since grown from four people to roughly 30, and Workday plans to continue investing in it. Gatto attributed that growth to what the team accomplished and to its work with LangChain. He also said LangSmith’s speed and maturity helped the group deliver value to customers. Those comments describe Workday’s experience; the account does not provide independent benchmarks comparing the platform with other enterprise agent systems.
For teams considering a similar approach, the notable design choice is the boundary between developer freedom and platform control. LangGraph supplies a flexible way to express agent workflows, while Workday’s SDK and internal services provide a common path to models, retrieval and other tools. The next measure to watch is whether the rapid staging path also supports reliable production operation as more teams build agents and depend on shared services.
Workday’s implementation draws on LangGraph, with LangSmith used for deployment and observability capabilities described at LangSmith Deployment and LangSmith Observability.