Twenty-three servers in one quarter, and security approved none
A coding agent that read a .env file, a stream whose consistency nobody wrote down, and three product managers who settled the number before the architecture.
How we Organise
The measure was agreed before the architecture, and that is what the team credits
Three principal product managers at Spotify published a joint account on 9 November. Their roles divide across a research layer, a shared recommendation model and the listener-facing surface. The one running the model names two levers: what it is taught, and how what it learned is checked. His team’s strongest result of the year came from a partner team that wanted a particular outcome. They agreed how success would be measured before they agreed what to build.
We tell a client that only a number settled in advance can be missed. One picked afterwards is a description rather than a commitment. Our practice puts it plainly:
A number agreed before a release can be missed; a number chosen after it describes what happened.
Spotify’s account is one firm on its own practice, with no comparison and no count of the times the sequencing was skipped. Its experience-layer half says evaluation is continuous rather than fixed, which may be the harder version of the same point. A measure settled first only helps while it still measures the same thing.
How we Build
A refund can be paid twice because two delays were given one name
An analysis published on 8 November takes apart real-time stacks that hold the same state in a stream processor and a serving store. It separates three failures that a single word hides. Export lag is the wait between a processor working out a value and the serving store receiving it. Replica lag sits inside the serving store. The third is a race where the deciding read and the update are not atomic, and a refund goes out twice. One read-modify-write across separate systems takes four or five network steps, by the author’s count. Where the computation runs on the state, it takes one.
We ask that a stream carry its own properties rather than leave a consumer to work them out:
Every stream this team publishes to or consumes from states its delivery, its ordering and its consistency.
Each of those three delays has a different repair. A consumer reading a value cannot tell which one it is subject to unless somebody wrote the properties down. The piece argues throughout for one vendor’s architecture. It is not a measured comparison, and says so itself, so what survives here is the distinction rather than any figure about cost.
How we Assure
Developers installed twenty-three MCP servers in one quarter and security approved none
An adviser to enterprise programmes wrote this up in CIO on 9 November. His argument is that coding-agent vendors belong in the supplier file, along with the model providers, plugin marketplaces and registries behind them. He reports a Fortune 100 bank whose agent took a .env file into its working context because the available restrictions were never configured. He reports a healthcare payer whose vendor swapped model providers mid-contract. No clause covered it. The seats had been bought on a corporate card, below the procurement threshold. At a specialty retailer, developers installed twenty-three MCP servers in a quarter and two could write to the CRM. He also notes that OpenAI told Cursor’s new owner on 28 August it means to stop supplying models, with 12 November proposed.
Our position is that an approval only ever covers the thing that is actually running:
Every approval this team relies on attaches to the version now running: since the last change to the model, the prompt, the tools, the corpus or the running state, a rollback, a restore or a replayed snapshot included, every check and every reading was taken again.
The test behind it reads two dates a team keeps about itself: its last reviewed change, and when its running state was last restored. A supplier that substitutes a model part-way through a contract changes what every approval was granted against, and it appears in neither date. This is not the first source to reach that statement. Trail of Bits made the same point in September, about an approval that outlives a change to the trusted execution environment beneath it. The account here is one adviser’s, the engagements are unnamed, and the wider figures in the piece come from reports we have not opened. Perhaps third-party risk process already covers these suppliers, and the failure is only in applying it. If you read it that way, write and tell us: transform@dromologue.ai.