AI
Engineering Foundations Came First
Nordstrom’s AI enablement team had engineering standards, sample applications, deployment pipelines, and a composable build system before MCP arrived. Ola Hungerford emphasized they did not get lucky. They invested in those foundations for years. The first MCP servers connected to the project tracker and internal wikis. They deployed those servers using the existing container platform instead of building AI-specific infrastructure. The MLOps team independently followed the same approach for ML workflow templates.
A Spectrum From Skills to Orchestration
Sandeep Bhat described a spectrum of AI solutions at Nordstrom. At one end are complex agent orchestrations with multi-step workflows and multiple tools. In the middle are dedicated MCP servers connected to internal systems. At the other end are skills and local experiments, including local MCP servers. Most use cases land in skills. That is by design. Nordstrom wants teams to experiment in the open, not in silos. The registry captures every server, skill, or agent with ownership, data access, and compliance status.
Registry and SDK Standardize the Platform
The registry uses the open-source MCP registry schema as its source of truth. Each entry includes enterprise metadata: support team, compliance status, data classification. Running an inner source model means anyone can submit a pull request. When something gains traction, it becomes a standard. Nordstrom also built an internal SDK, a lightweight wrapper around the MCP SDK. Any application using it gets OAuth and tool call observability for free. Teams focus on business logic; the platform handles the rest.
Feedback Loops Keep Documentation Alive
Ola Hungerford stressed that human feedback loops are the hardest part. Regular office hours turn into a feedback loop for standards, guides, and sample applications. They expose duplicate efforts, like multiple teams building incident response agents. Sandeep Bhat described a system that monitors Slack support threads, waits for a thread to cool off, determines if something was missing from docs or if a bot gave a wrong answer, then creates a pull request to update the documentation. That way knowledge trapped in Slack becomes agent context.
Notable Quotes
Uh MCP servers are services. Ola Hungerford · ▶ Watch (10:37)
if our agent disappeared tomorrow, would the knowledge survive? Sandeep Bhat · ▶ Watch (6:57)
experimenting in the open is another big part of what we encourage. Ola Hungerford · ▶ Watch (10:51)
Key Takeaways
- Nordstrom reused existing container platform for MCP servers instead of building new infrastructure.
- An internal SDK provides OAuth and observability, letting teams focus on business logic.
- A Slack-based system automatically updates documentation from support conversations.
About the Speaker(s)
Ola Hungerford is a Principal Engineer at Nordstrom and a maintainer and community moderator for the Model Context Protocol. She leads AI enablement initiatives while contributing to MCP’s specification, developer tooling, documentation, and community governance.
Sandeep Bhat works on the AI Enablement team at Nordstrom, focusing on platform engineering for MCP and AI agents. His past background in security grounds his decision-making, allowing him to balance safe architecture with a passion for AI productivity. He builds infrastructure that empowers internal teams.