Dropping a file into `~/.claude/agents/` gets you a definition, not a dispatcher. The tally in the dev.to post shows where the work went instead: 19 calls to `Explore`, 3 to `general-purpose`, 1 to `code-reviewer`, no errors [4]. Twenty-two of the twenty-three logged calls went to the two generic routes [1]. The author's own explanation is the load-bearing sentence: unless an agent is named, the model takes the generic route, and a description reading "MUST BE USED for all code changes" is text inside a file rather than routing logic [7]. Selection happens on the calling side, in the prompt or in hook code [8].
Then look at the file list. Eight files, and one of them is `INDEX.md` [2]. The zero-call report lists INDEX as an agent that was never invoked [3]. A table of contents that has never been called is behaving correctly. So the reported 87.5% [5] is really 6 of 7 plausible agent definitions, or 85.7% [2]. Same direction. But a counter that cannot distinguish a catalog file from an agent will also miscount the day someone adds a README.
The logging itself needs a second look. The log holds 682 records spanning May 28 to August 30, 2026 [6], and 23 of them fall in the audited 30 days [4]. That leaves 659 records across the preceding 65 days, about 10.1 per day, against 0.77 per day in the window under audit [3], a thirteen-fold gap. Either the author's usage collapsed, or the stop hook stopped catching some class of invocation, and the post does not separate the two. For "zero calls" to mean "never ran", the hook has to fire on every subagent completion, including the ones you triggered by name.
The token argument is the thinnest part of the case. The author asserts that definitions are injected into the system prompt on every request and points at `architect.md` running past 220 lines [9][10]. Lines are not tokens, and no measurement of the injected catalog appears in the post. The direction is plausible; the magnitude is unstated, which makes it a reason to measure rather than a reason to delete.
What I would copy is layer one: a stop hook that appends one JSONL record per invocation with timestamp, session id and subagent name [11]. That is the cheapest instrument in the stack, and almost nobody builds it, because agent definitions feel like configuration instead of code paths. He set out to measure what he already believed, and the log happened to back him up.
The 87.5% figure only transfers to your setup if the hook logs every subagent call, your prompts rarely name an agent explicitly, and you were expecting autonomous selection to happen in the first place. If your prompts do name agents, the counts measure your prompts, not the router [8].
