Presenters

Source

🤖 Beyond the UI: Welcome to the Age of the Agentic Backstage

In the world of internal developer portals, we’ve spent years obsessing over the user interface. We’ve built beautiful dashboards, complex plugins, and intuitive search bars to help humans navigate the software wilderness. But according to Sam Nixon from Roadie, the tide has turned.

We are no longer building just for humans. We are building for agents.

At a recent talk, Sam shared a staggering statistic from the Roadie platform: the ratio of agent interactions to human interactions has hit 100 to 1. While human usage remains static, agent activity is exploding. This shift demands a fundamental rethinking of how we build and maintain our Backstage catalogs.

📈 The 100:1 Reality Shift

Backstage began as a tool to help engineers find information. However, the data reveals a new trajectory for 2025 and 2026.

  • 2021-2023: The focus was on catalog completeness and reducing YAML toil.
  • 2024: The rise of RAG (Retrieval-Augmented Generation) frameworks.
  • 2025/2026: Agent interactions have crossed the threshold, exponentially outperforming human logins.

The lesson? Your UI is no longer the most important part of your portal. If 100 agents are calling your MCP (Model Context Protocol) server for every one human looking at a table, your priority must shift toward machine-readable data.

🧩 The “Assistant” Problem: Why Metadata Matters

Roadie experimented with a Claude 4 powered assistant to help resolve incidents. They quickly hit a wall. When an alert reported that the payment service was down, the agent failed because the metadata was poor:

  • Vague Ownership: The owner was set to a generic Engineering Department with 134 members.
  • Missing Links: No PagerDuty annotation was set.
  • Generic Docs: The runbook was too broad to be actionable.

For an agent to be effective, it doesn’t just need an LLM; it needs a comprehensive, fresh, and enriched graph of your software topology. 🌐

🏗️ Building the Agentic Foundation

Sam outlines three pillars required to transform Backstage into an agent-friendly ecosystem:

1. A Fresh Metadata Graph 📊

Agents require real-time data. Relying on manually updated YAML files is a recipe for failure.

  • The Solution: Use Providers and Processors to ingest data dynamically from GitHub, AWS, Kubernetes, Data Dog, and PagerDuty.
  • Pro Tip: Sam recreated most of these providers using Claude in just one day—one specific experiment took only 38 minutes. You can build these tools faster than ever.

2. Agentic Relationship Building 🔗

Nodes in a graph are useless without edges. Agents need to know how an AWS S3 bucket relates to a specific React component.

  • Tools: Use inexpensive models like Claude 3.5 Sonnet or Qwen to draw these connections algorithmically.
  • The Tradeoff: Avoid strict hierarchies and rigid kinds (component, system, domain). Agents reason better with descriptive, declarative names rather than constrained schemas.

3. Machine-Accessible Formats 📡

Agents “hate” tables. They need data delivered via:

  • MCP Servers
  • CLIs
  • APIs
  • Vector Databases
  • Minimal Token Counts: Optimize your reports to provide the most context with the fewest tokens to save on costs and latency.

⚡ From Reasoning to Action: The Scaffolder’s New Role

The “pot of gold” at the end of the rainbow is moving from an agent that knows things to an agent that does things. 🦾

By combining the Software Catalog with the Backstage Scaffolder, you enable agents to:

  • Investigate and resolve incidents autonomously.
  • Run migration refactors.
  • Rotate secrets or deploy new services.
  • Create Slack war rooms and Jira tickets automatically.

Because Backstage has a robust permission system, you can wrap these tool calls in layers of security, ensuring agents only act within their authorized boundaries.

🎯 Real-World Results: The 80% Win

Does this actually work? Sam shared the impact of these agentic workflows at Roadie:

“We find that we can resolve about 80% of our support requests and on-call alerts via this system without the need for an engineer to actually do anything.”

While the remaining 20% of complex issues still require human expertise, the order-of-magnitude shift in speed and efficiency is undeniable.

❓ Q&A Highlights

Audience Member: Do you have metrics on the implementation, like resolution time or the amount of incidents made easier for developers?

Sam Nixon: We see about an 80% faster resolution for the incidents handled by this process. For us, the support engineer gets paged, drops into a Slack channel, and reads the summary generated by the agent—which has already found the runbook, identified the on-call person, and linked the relevant Prometheus queries. The goal is to eventually automate the triage point entirely so humans are only involved downstream.

🚀 Final Thoughts

The future of Backstage isn’t just a better catalog for developers; it’s an operating system for AI agents. By focusing on fresh metadata, rich relationships, and actionable tools, you can transform your developer portal from a static directory into an autonomous engine of productivity.

Are you ready to stop building for users and start building for agents? ✨

Appendix