Backstage’s journey from spreadsheets to global IDP standard | Spotify’s Tyson Singer
The latest in Gastown, the death of the IC, and Claude Code for the rest of us.
Before Backstage became the industry standard for developer portals, Spotify’s engineers relied on spreadsheets to navigate their massive microservices ecosystem.
Tyson Singer, Spotify’s Head of Technology and Platforms, joins us to trace the evolution of their internal developer experience from a necessity for order into the open-source giant Backstage and its new SaaS evolution, Portal. We dig into how they use golden paths to align autonomous squads and how their new AI Knowledge Assistant (AiKA) reduced internal support tickets by nearly 50% while protecting developer flow. Finally, Tyson shares his philosophy on sustainable innovation, explaining how to train an engineering organization to run a marathon at a sprinter’s pace.
1. Cowork: Agentic workflows for the rest of us
Anthropic just launched Cowork, a feature for subscribers that lets the AI access local folders to read, edit, or create files. This moves Claude from a chatbot to an autonomous worker capable of executing multi-step tasks. It effectively takes the complex agentic loops developers are building and democratizes them for general knowledge work. The goal isn’t just efficiency. It is about giving the AI the keys to the file system so it can stop offering advice and start committing actual changes.
Read: Cowork: Claude Code for the rest of your work
2. Gas Town Emergency User Manual
Gas Town has quickly evolved from a niche experiment to the inescapable blueprint for agent orchestration. It is a new open-source system from Steve Yegge designed to manage massive numbers of AI coding agents. Think of it as “Kubernetes for agents.” The core breakthrough here is the Molecular Expression of Work (MEOW). This concept breaks tasks down into such granular steps that ephemeral agents (called “Polecats”) can pick them up, execute them, and hand them off without losing context. It is experimental, expensive, and complex, but it represents the shift from single-agent coding to agent assembly lines.
Read: Gas Town Emergency User Manual
3. Loom: The orchestrator you aren’t supposed to use
Geoffrey Huntley has open-sourced Loom, an orchestrator based on the viral Ralph Wiggum technique. Ralph is essentially a bash loop that feeds an AI’s errors back into itself until it stumbles onto the right answer. It is brute force meets persistence. Huntley is explicit: “If your name is not Geoffrey Huntley then do not use Loom.” This warning is code for studying it, understanding the failure modes, and building your own.
Read: ghuntley/loom
4. AI Killed the Individual Contributor
The role of the software engineer is shifting from lone gunslinger to orchestra conductor. A new article argues that as AI tools mature, the value of an IC is moving from lines of code written to lines of judgment applied. A great day isn’t laying 350 bricks yourself anymore; it’s unblocking the agents that laid 350. The IC isn’t disappearing, but they are evolving into managers of digital workers. The skillset is pivoting from syntax to system design, priority setting, and curation.
Read: AI Killed the Individual Contributor
5. The URL shortener that makes your links look suspicious
CreepyLink is a lesson in doing the exact opposite of what the internet typically strives for, which is building trust. It is a URL shortener designed to make your links look as sketchy as possible. By deliberately looking untrustworthy, it pokes fun at our assumptions about security and aesthetic safety. It serves as a playful reminder that “trust” is often just a UI choice.
Read: CreepyLink
6. ICYMI: What Meta’s “diffs per developer” metric actually reveals
When James Everingham returned to Meta for his second tour, he encountered a metric that dominated leadership updates: DDM (diffs per developer per month). Leadership wanted to return to pre-COVID levels, but the metric ignored the reality that the work had gotten “heavier” due to a massive codebase and stricter compliance. James compares it to running a five-minute mile on Earth versus doing the same on Jupiter—same runner, different gravity. This serves as a lesson in Goodhart’s Law: when a simple metric becomes a target, teams optimize for motion rather than outcomes, slicing work thin to satisfy the chart rather than solving the gnarly problems that move the needle.
Read: What Meta’s diffs per developer metric revealed about engineering at scale
7. Join me at The Atlantic x Infactory Hackathon
At the end of the month, I am heading to Palo Alto to help build the future of news with our friends at Infactory and the team at The Atlantic. Infactory’s Brooke Hartley Moy (a Dev Interrupted alum) and John Kanalakis are taking The Atlantic’s massive digitized archive and challenging us to ask how AI can build trusted news services that serve readers. It is exactly the kind of bold experimentation journalism needs right now. The judges include Nicholas Thompson, CEO of The Atlantic, and Ken Kocienda, inventor of autocorrect and recent Dev Interrupted guest. I will be there hacking away, so if you are in the Bay Area, grab an RSVP and come say hi.
Read: Building future AI news experiences
Fascinating trajectory from spreadsheets to Portal. The 50% reduction in support tickets through AiKA is wild, but what caught me was the golden paths mention. At my last gig we tried similar developer portal patterns but struggled with keeping documentation fresh enough that people actully trusted it over just asking someone. The shift Singer describes from IC writing code to orchestrating agents feels like the next version of that same documentation trust problem, just at a diffrent abstraction layer.