On a recent AI capacity buildout, I scheduled a control plane node into an early cutover wave because the 2 week traffic snapshot I had of it looked effectively isolated. It was not. It was a monthly capacity reconciliation job. On the 21st it woke up, half the reconciliation now crossed a data center boundary, and pipeline latency moved from roughly 10 minutes to close to 60.

The graph was not wrong. My observation window was. We had used a 2 week capture window as the source of truth for a job that only ran once a month. The failure mode I care about is a reasoning model that looks right on the evidence it was shown, and quietly wrong on the evidence it was not. A single model handed all of this evidence will smooth the contradictions between it and produce a confident answer. A small team of specialist agents, with a coordinator that names contradictions instead of averaging them, does not.

The work at this scale is not one question. It is dozens of questions that share very little context with each other. A single strong reasoning model failed in 2 ways: it averaged away contradictions or dropped tasks. The subagent pattern that Codex and Claude Code both ship solves the second failure directly, because each specialist only ever sees the evidence for its own question. Each specialist returns a typed finding, not a paragraph. Claim, evidence, confidence, freshness, and one of 3 verdicts: ok, warn, or hold.