Here is the oldest problem in scaling engineering: a good decision, made once, is lost a thousand times. A team works out the right way to authenticate, to lay out a service, to ship to production — and the next team, and the next, re-litigates the same questions from scratch, usually worse. Multiply that across an organisation and you get sprawl: a hundred slightly different ways to do the same thing, none of them the way you’d choose. The golden path is the answer to that problem. AI agents have just made it non-optional.
What a golden path actually is
Spotify, which coined the term, defines a golden path as “the opinionated and supported way to build” something — a paved road through the organisation’s tools and standards. The crucial word is supported: a golden path is a default, not a mandate. At Spotify it is explicitly optional — you may leave it, but you forfeit the support that comes with staying on it. It was introduced to solve ecosystem fragmentation and to end “rumour-driven development”, where teams guess at the right approach from hallway conversations. And it is sanctioned centrally: a Technical Architecture Group decides which frameworks the path endorses, so the path carries the organisation’s considered judgement, not one team’s preference.
A golden path is intent, compiled
Strip it back and a golden path is a mechanism for preserving intent. The architecture decisions, the security and operational requirements, the conventions a senior team agreed once — a golden path encodes them into the default route, so they survive across teams without being re-argued every time. It is how a decision made once survives a thousand repetitions. That is the same job our Decision Architecture and Intent Translation work does at the level of a single programme; the golden path does it at the level of the whole estate.
A golden path is an architecture decision that only has to be made once. Everything after that is inheritance.
Templates are how the path travels
Intent that lives in a document gets ignored; intent baked into the tooling gets followed. Service templates are how a golden path actually travels — scaffolding (Backstage’s Software Templates being the best-known example) that stamps out a new service already wired for the current standards, so a team inherits them by default rather than reading about them. The catch is drift: once a service is instantiated from a template, it stops tracking the template, because feeding improvements back upstream is tedious and rarely done. A golden path is only ever as current as its templates — and as honest as the gap between what the template produces and what the estate actually looks like.
The golden cage
Push opinionation too far and the paved road becomes a golden cage: so constrained that capable teams can’t do legitimate things, with “ten-foot drops either side” if they step off. The deeper failure is subtler — a path that is mandated but painful. Team Topologies’ corrective is to treat the platform as a product and its developers as customers who choose it: you build the Thinnest Viable Platform — no thicker than it needs to be, sometimes as thin as a wiki page — and you measure whether people actually adopt it, the way you’d measure any product. The principle DORA’s research keeps returning to is the same: self-service golden paths earn their keep by reducing cognitive load. A path that adds load to enforce conformance has missed the point.
A paved road only works if it is genuinely the easiest road. Mandate a painful one and people quietly pave their own.
Now the path serves non-human contributors
The newest travellers on the golden path are AI agents — and they need the same intent, delivered in a form they can use. Three practices are converging. A reference application — a live, compilable codebase rather than a folder of stale snippets — gives an agent worked examples to keep its output architecturally consistent. An AGENTS.md file (an open, vendor-neutral standard, stewarded under the Linux Foundation since 2025 and adopted across tens of thousands of projects) puts agent-specific instructions where the agent looks, distinct from the human README. And a Model Context Protocol server can expose that reference code and its commit history to agents, so they can detect drift and propose repairs against the design intent. Bake those into the service template and the golden path becomes the distribution mechanism for agent context: every new repository inherits the latest human and machine guidance by default.
In the agent era, the golden path is the only scalable way to give every contributor — human or agent — the same intent.
What to measure
Treat the golden path as a product and watch the signals a product owner would: the share of services actually on the path (adoption, not mandate); time-to-first-deploy on the path versus off it; how far instantiated services have drifted from the current template; and how current the template itself is. We offer these as the questions to track rather than industry benchmarks — the field hasn’t settled on agreed thresholds yet. The single number that matters most is adoption: a path people and agents choose is working; a path you have to enforce is already failing.
This is where intent stops being a document and becomes the default. A golden path is the most concrete form delivery architecture takes — the place an organisation’s considered judgement is compiled into the route of least resistance, and handed, unchanged, to whoever (or whatever) builds next. That was always the promise of platform engineering. Agents simply made it the difference between scaling your standards and scaling your sprawl.