Here's the answer, and it's smaller than it sounds: a subagent is a specialist with its own instructions and its own isolated workspace, defined in a simple file, a name, a description of when it should be used, and its standing orders, with controls for which tools it may touch and which model it runs on. The system invokes the right specialist automatically when a job matches its description. Hiring is literally writing the file.
Hire where your assignment library shows repetition, and four hires cover most operations: the researcher, your receipts-required brief as a permanent employee that only reports what it verified. The reviewer, your acceptance criteria as an adversarial checker, because writer and reviewer should be separate. The producer, your content batch or report format as a specialist. And the auditor, checking work against your governance page. Scope each to its narrowest useful toolset, a researcher needs web access, not file-deletion powers, and the cheapest model that does its job well.
Then probation, exactly like a human hire: the first four runs reviewed in full, corrections written into the file rather than left in chat, permissions widened only after clean runs. And one rule protects the whole roster: don't hire what you haven't done well manually, because the file encodes your standard, and a guessed standard gets repeated perfectly forever.
Parallel work has one test: if two jobs share no inputs and no outputs, they can run at once. And the coherence rules keep ten workers on one truth: shared context lives in files all workers read, and every parallel output merges through one desk before anything ships.
The full staff model is Volume 3 of our Advanced Operator series.
Hire in Markdown. Probation always. One truth, one desk.