Building AI Agents on Workday Extend: A Developer's Guide for 2026 (2026)

Workday's platform has evolved from a closed ERP system into something that increasingly resembles a developer platform. With the Build Platform (combining Extend, Flowise, and Data Cloud), developers can now create AI agents that operate within Workday's security model while leveraging external AI capabilities. This guide covers what developers need to know to build production-quality AI agents on Workday in 2026.

The Build Platform Stack

Workday's Build Platform is the unified development environment that brings together three previously separate capabilities into a coherent stack for building AI-powered applications.

• Workday Extend — the application framework for building custom apps that run inside Workday's UI and access Workday data through sanctioned APIs
• Flowise — Workday's visual AI orchestration layer for designing agent workflows without writing low-level prompt engineering code
• Data Cloud — the data access layer that exposes Workday data through Apache Iceberg tables for analytical workloads and AI model consumption

Together, these three layers give developers the ability to build agents that can read Workday data, reason about it using external LLMs, take actions within Workday, and learn from the outcomes — all within Workday's tenant security boundary.

Agent Gateway: MCP and A2A Protocols

The Agent Gateway is Workday's interface layer for connecting external AI agents to Workday's internal systems. It implements two emerging standards that developers should understand.

Model Context Protocol (MCP) provides a standardized way for AI agents to discover and invoke Workday capabilities. Rather than hardcoding API calls, agents use MCP to dynamically learn what actions are available in a given tenant based on the user's security profile. This means the same agent code can work across tenants with different configurations.

Agent-to-Agent (A2A) Protocol enables Workday's native agents to communicate with custom agents built on Extend. This is bidirectional — your custom agent can delegate subtasks to Workday's built-in agents (like the Financial Close Agent or Recruiting Agent), and Workday's agents can invoke your custom agent when they encounter scenarios you have registered to handle.

• MCP discovery endpoints expose available Workday business processes as agent-callable tools
• A2A message passing uses structured schemas for inter-agent communication
• Security context flows through both protocols — agents inherit the invoking user's permissions
• Audit logging captures all agent-to-agent interactions for compliance review

Development Workflow

Building an AI agent on Workday Extend follows a specific workflow that differs from traditional web application development. Understanding this flow prevents common mistakes.

First, define the agent's purpose and scope within Workday's domain model. Agents that try to do too much across too many functional areas become impossible to test and maintain. The best agents solve a specific problem within a bounded context — for example, an agent that helps managers understand their team's compensation positioning relative to market data.

• Design the agent's tool set — what Workday APIs will it need to call?
• Define the data access patterns — will it use real-time APIs or Data Cloud analytical queries?
• Map the security requirements — what roles and permissions must users have?
• Build the Flowise workflow — define the reasoning chain and decision points
• Implement the Extend UI — create the conversational or dashboard interface
• Register with Agent Gateway — make the agent discoverable via MCP and A2A
• Test in sandbox — validate against realistic data volumes and edge cases

Data Access Patterns

Agents need data, and Workday offers two distinct patterns for agent data consumption. Choosing the right one depends on latency requirements and data freshness needs.

For real-time queries against transactional data, use Workday's REST and SOAP APIs through the standard Extend integration layer. This gives you current-state data with full security trimming. The tradeoff is query complexity limitations and rate limiting at scale.

For analytical queries that span large datasets or historical periods, use Data Cloud's Apache Iceberg tables. These provide SQL access to denormalized views of Workday data that update on a near-real-time basis. Agents performing trend analysis, anomaly detection, or benchmarking should use this path.

Security Considerations

Every developer building AI agents on Workday must internalize one principle: agents act on behalf of users, not as system-level processes. This means the agent's capabilities are bounded by the invoking user's security profile. An agent cannot access data or perform actions that the user themselves cannot access or perform.

This has implications for agent design. If your agent needs to aggregate data across organizational boundaries — for example, comparing compensation across all departments — it will only function correctly when invoked by users with broad visibility. Design your agent's error handling to gracefully degrade when security trimming restricts the available data.

What Developers Should Learn Now

The Build Platform is evolving rapidly, and developers who invest in the fundamentals now will be positioned to build production agents as the platform matures through 2026 and into 2027.

• Learn Workday's object model — understanding Workers, Organizations, Positions, and their relationships is prerequisite knowledge
• Study MCP specification — the protocol is not Workday-specific and is gaining adoption across enterprise platforms
• Practice with Flowise — visual orchestration is different from code-first development but produces more maintainable agent logic
• Understand Workday security — ISSG-based security is unique to Workday and fundamentally shapes what agents can do

Building AI agents on Workday Extend is a genuinely new capability that did not exist in this form even a year ago. Developers who learn the platform now — while the ecosystem is still forming — will have a significant advantage as organizations begin deploying custom agents at scale.