Workday says its shared AI platform cut agent prototyping from months to under a day

Workday teams can now get an AI agent prototype running in a test environment in under a day, rather than four to five months, according to a LangChain case study published November 10. The account describes an internal platform that puts authentication, security and deployment infrastructure underneath the agent-building work, instead of leaving each team to assemble them.

The platform was built for Workday engineers developing agents on the company’s own infrastructure. In 2025, leadership set a goal of having agents in production by the end of 2026. A four-person team initially had to cover a scalable agent server, data access controls, shared evaluation tools and a software development kit.

Flexible workflows on a common foundation

Workday separated the way engineers design an agent from the infrastructure that runs it. LangGraph’s graph-based model lets engineers build simple agent loops or more complex workflows with predetermined steps, rather than fit every agent into a fixed sequence of conversations or personas.

• LangGraph handles orchestration: coordinating the steps that make up an agent’s work.
• LangSmith provides tracing, monitoring and evaluations: tools for inspecting what agents do and assessing their results.

The team built a Python software development kit—a reusable set of programming tools—around LangChain’s building blocks. Alex Gatto, a senior machine learning engineer, calls it “Workday LangChain.” It routes those building blocks through internal services, including a gateway for language models and a retrieval service for supplying relevant information to them.

Permissions are part of the platform

Workday’s internal services must meet audit, compliance and data retention requirements before entering production. Its Agent System of Record controls which application programming interfaces, or APIs, agents may access. The platform team built the backend server and evaluation infrastructure beneath that governance layer.

Gatto says support for the A2A and MCP protocols was already in production when leadership asked about it. Both came with LangChain’s server implementations, so Workday’s platform team did not need to add that support separately.

The platform also connected previously isolated internal services, including model gateways, retrieval tools, recommendation engines and prompt services. Gatto describes it as a unified interface that makes those tools accessible beyond the teams that originally built them.

The case study reports faster prototyping while the Agent Platform team grew from four people to roughly 30, with further investment planned. Gatto credits the LangChain partnership and the team’s accomplishments for that growth.