It support · Production

Expedia Group reimagines platform engineering to serve AI agents alongside humans

The problem

Expedia Group's platform was designed for human engineers — microservices, SDKs, and UIs optimized for human ergonomics — and was not ready to support agents as a distinct user group. Proprietary abstractions on top of standard tech made the platform unfamiliar to agents trained on the open ecosystem.

First attempt

When asked to perform platform tasks without proper agent interfaces, agents routed around the platform by logging into web UIs through browser automation — inspecting cookies and session state — producing brittle, unreliable operations that the platform team explicitly wanted to avoid.

Workflow diagram · grounded in source
1
Agent receives platform task
trigger
“People have asked agents like Claude or others to "watch my pipeline" even when there is no standardized API or integration to do that”
2
Agent routes around missing interface
ai_action
“The agent figures out how to log in through the browser, inspects cookies and session state, then navigates the UI as a human would, just faster and more persistently”
3
Tarmac CLI exposes platform operations
integration
“a new CLI we call Tarmac. It's an agent‑centric interface that exposes CI/CD operations, repository management, workload monitoring in our Kubernetes clusters, and log exploration and related workflows”
4
MCP servers enable capability discovery
ai_action
“We've been aggressively rolling out Model Context Protocol (MCP) servers and a registry for our core platform capabilities so agents can: Discover what's possible. Call well‑defined operations. Get structured responses instead of scrapin…”
5
Agent assembles internal apps via Koda
ai_action
“Koda lives in a monorepo dedicated to agent‑built internal apps and focuses on context engineering, not on building yet another monolithic platform. It lets agents discover existing APIs, simplify authentication flows on behalf of users,…”
6
Observed friction feeds platform improvement
feedback_loop
“When we see agents hacking around UIs, we respond by adding better agent interfaces. When we see complex, brittle flows, we respond by simplifying contracts and tightening guarantees. When we see friction, we respond by improving ergonom…”
Reported outcome

Expedia Group began shipping agent-native infrastructure including the Tarmac CLI (covering CI/CD, Kubernetes, and log exploration), MCP servers with a capability registry, markdown-based agent skills packaging tribal knowledge, and the Koda internal app platform.
A 'no-coding-allowed' Ralphathon hackathon confirmed the hypothesis that structured experimentation — not just instruction — is required to shift engineer workflows.

Reported metrics
Engineering work automatable by agentsA huge amount of what we call "engineering work" can now be done by agents
Software production gainsclear gains in how much software can be produced in a given amount of time
Agent operational cost and efficiencyagents spend less time fumbling around and more time doing useful work. And the cost profile of that work improves
Reported stack
TarmacModel Context Protocol (MCP)ClaudeKodaBackstageKubernetes
Source
https://medium.com/expedia-group-tech/reimagining-platform-engineering-for-an-agentic-future-03e3f378a190
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Expedia Group began shipping agent-native infrastructure including the Tarmac CLI (covering CI/CD, Kubernetes, and log exploration), MCP servers with a capability registry, markdown-based agent skills packaging tribal…

What tools did this team use?

Tarmac, Model Context Protocol (MCP), Claude, Koda, Backstage, Kubernetes.

What results were reported?

Engineering work automatable by agents: A huge amount of what we call "engineering work" can now be done by agents; Software production gains: clear gains in how much software can be produced in a given amount of time; Agent operational cost and efficiency: agents spend less time fumbling around and more time doing useful work. And the cost profile of that work improves (source-reported, not independently verified).

What failed first in this deployment?

When asked to perform platform tasks without proper agent interfaces, agents routed around the platform by logging into web UIs through browser automation — inspecting cookies and session state — producing brittle, un…

How is this it support AI workflow structured?

Agent receives platform task → Agent routes around missing interface → Tarmac CLI exposes platform operations → MCP servers enable capability discovery → Agent assembles internal apps via Koda → Observed friction feeds platform improvement.

Related it support cases