Foundations Before Floors: What to Standardize in Agent Platforms
- Source
- writer.com
- Author
- WRITER
- Date

Context
WRITER builds its platform on a rule its Enterprise Platform lead, Raafat Zarka, sets out in an engineering post: share the parts enterprises will always need, and put only a stable interface in front of the parts that keep changing. He decides with three questions: how durable the need is, how expensive it is when teams diverge, and whether this is where WRITER differentiates. Identity, authorization, encryption, audit trails, usage metering and cost attribution pass, because customers always need them and inconsistency there is costly. Model providers, agent frameworks and protocols change every few months, so for those WRITER standardizes the seam, the interface other systems plug into, and lets what sits behind it change. Its gateway is the example: every connector, WRITER-built or custom, looks the same to an agent and passes the same schema and prompt-injection checks before it enters the registry.
The post weighs both directions with WRITER's own decisions. Logging for model calls, as the announcement puts it, shipped in March inside a single service and did not cover agent runs, so by September it had been rebuilt on the company's data platform, where any event more than one team depends on travels one path: published against a schema, validated, deduplicated, encrypted where sensitive and loaded into a warehouse. Spend alerts went the other way. Customers wanted an early warning, so WRITER shipped notifications first, with a scheduled job checking totals every minute instead of the infrastructure it lacked. That shortcut still ran through the data platform and used the same usage events as billing, which is why Zarka says the planned hard spend limits will build on it rather than replace it.
The call WRITER got wrong was a standard nobody made the default. The team agreed that every user-facing page belongs in one frontend monorepo, but pages kept landing in the older app because that was the quickest route, and the two drifted onto different design-system versions until the same menu showed different colors depending on which app served it. New pages now go into the monorepo unless someone argues otherwise before starting. Zarka adds that agents copy the nearest example even more faithfully than people do, and in a later section, citing OpenAI's similar conclusions, he says WRITER now holds every repository its teams own to an agent-readiness standard: a short root AGENTS.md pointing to deeper docs, explicit must-not rules, and requirements enforced in CI rather than only written down.
- WRITER decides what to standardize by asking how durable the need is, how costly divergence is, and whether it differentiates there. Identity, audit trails, encryption and cost attribution get shared foundations; for fast-changing models it standardizes only the interface.
- To ship spend alerts quickly, WRITER used a once-a-minute scheduled job instead of event-driven streaming it lacked, but fed it from the same usage events billing uses. Because the shortcut sits on the main platform, planned hard spend limits will build on it, not replace it.
- At WRITER, any event more than one team depends on travels one schema-defined, validated, deduplicated path. Model-call logging first built inside a single service missed agent runs when it shipped in March and had to be rebuilt on the shared data platform by September.
- WRITER agreed all user-facing pages belonged in one frontend monorepo but never made it the default path, so pages kept landing in the older app, causing design-system drift and missing analytics. The author's lesson: a standard that isn't the default is a suggestion.
- WRITER holds repositories to an agent-readiness standard: a short root AGENTS.md pointing to deeper docs, must-not rules enforced in CI, and recurring tasks packaged as skills with 'done when' checks. When an agent errs, the team fixes the , not just the PR.
- AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
- MCP — The Model Context Protocol — an open standard that lets any AI assistant plug into any tool or data source without custom integration code.
- streaming — Sending a model's response token by token as it is generated, so the reader sees text immediately instead of waiting for the whole answer.
- context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
Checking sign-in…
Loading comments…






