Cerebras published an unusually honest account of how they built their internal knowledge base. It answers fifteen thousand questions a day, three months after launch, and the consumers are not just employees — automations and agents query it too. The engineering is genuinely good and unusually concrete: threads distilled into searchable form before embedding, four retrieval signals fused because no single scorer can be trusted alone, age decay because Slack answers expire. They name three pillars: a platform for collecting data, a platform for querying it, and a layer that enforces authentication and authorization with auditing. The first two pillars get several thousand words. The third gets one sentence.
That asymmetry is the interesting part, because it is about to be replicated everywhere. When a human searches a knowledge base, weak retrieval is an annoyance — the human notices the answer smells stale, cross-checks, asks a colleague. When an agent consumes the same knowledge base as ground truth, retrieval quality stops being a product metric and becomes a control. The agent does not have the human’s instinct for staleness. It takes what the retrieval layer returns and acts on it — writes the code, files the response, routes the escalation. Whatever the index believes, downstream systems now believe.
Look at what the engineering itself concedes. Cerebras built age decay into their ranking because two threads can answer the same question and the six-month-old one may describe infrastructure that no longer exists. That is a correctness control, built by engineers because bad answers annoyed people. But nobody is testing it the way you would test a control: what is the acceptable staleness rate, who accepts it, and how would you detect drift? The published evidence that the system works is usage volume. Fifteen thousand questions a day tells you people find it useful. It tells you nothing about how often it is confidently wrong, and confident wrongness is precisely the failure mode that matters once agents are the consumers.
Permissions are the second gap. A retrieval layer that spans Slack, code, documents, and personnel expertise — their who_knows tool returns people — is a need-to-know collapse waiting to happen. Per-source access control that made sense when humans queried one system at a time composes differently when a single agent holds grants across all of them and a single question fans out in parallel. Authorization needs to be asserted per query, at retrieval time, against the human on whose behalf the agent is acting. Cerebras may well do this. The point is that the essay — the honest, detailed, engineering-proud essay — did not consider it worth describing, and the hundred enterprises about to copy the architecture will copy the emphasis too.
None of this is a criticism of Cerebras. They solved the hard engineering problem and published real numbers and real thresholds, which almost nobody does. It is an observation about sequencing: the agent-facing layer becomes control infrastructure before anyone names it a control. The same thing is happening with agents that gate code review and incident escalation — the gate gets assured never, because it arrived as a productivity feature. Retrieval is following the same path. If your organisation is building a knowledge base that agents will consume, the assurance work is not a later phase. Retrieval quality needs an evaluation protocol with an acceptance bar, not a usage dashboard. Freshness needs an owner, because expiry is a property of knowledge, not a ranking heuristic. And authorization needs to hold at query time for the person behind the agent, not the service account in front of it. The engineering essay for that layer is the one nobody has written yet.