The Subagent Roster

# The Subagent Roster

Twenty-four specialists on this machine, and not one of them can ask you a question.

That sentence is the whole thing. A subagent either finishes or it fails. There is no path where it
comes back mid-run and says *"which repo did you mean?"* — so a vague brief is not a slow run, it is
a wasted one. The full cost, spent, with nothing usable at the end.

This is the specific way delegating to an agent fails differently from delegating to a person. The
person would have asked.

## Why you want the roster at all

The failure it prevents is ordinary and expensive: you try to audit nine files yourself, burn the
whole context window on the file dumps, and answer from the half you can still remember. Nothing
errors. You just quietly get a worse answer than you would have got from the first three files
alone.

A subagent reads those files somewhere other than your main thread and hands back a conclusion.
The reading happens; the dumps never touch your window. That is the entire mechanism, and it is why
"add more agents" makes every future run cheaper rather than more complicated.

Because of the fact at the top. If it cannot ask, the brief has to carry everything up front. Four
sections, and a brief that carries all four never needs a question:

1. Constants — baked, do not discover. Paths, names, thresholds, the exact file. Anything the
agent would otherwise have to go find is a place it can go wrong silently.

2. Do exactly this. The steps, in order. Not the goal — the goal is what you'd give a person who
can ask follow-ups.

3. Hard rules. What it may never do, stated as prohibitions rather than preferences. "Never
write outside this directory" survives a weaker model; "try to be careful" does not.

4. Return — ONLY this JSON, nothing else. A fixed return shape is what makes the result
checkable without reading the prose. It also means a partial failure looks like a schema violation
instead of a confident paragraph.

A specialist that guesses is worse than no specialist.

If you'd done the work yourself, you'd have noticed your own uncertainty. You will not notice
somebody else's — especially when it arrives in the right format, on time, in the shape you asked
for. That's the trade you accept when you delegate to something that can't ask, and it's the reason
this kit spends most of its pages on the brief rather than on the roster.

- SKILL.md — the picker (when a job wants a subagent at all) and the spec method
- references/agent-template.md — the house shape: the four sections above, as a file
- agents/ — three working definitions you can drop in: verifier.md, web-claims-verifier.md,
web-researcher.md

Only three, and that's deliberate: the rest of my roster is wired to a two-machine setup and would
not run for you. Three that work beat twenty-four that don't.

Install in about a minute: drop the agents/*.md into ~/.claude/agents/ and keep the template
open the next time you write a brief.

Comment ROSTER on the video and I'll send it over — though nothing is gated on that. It's at
sharpendaily.co/kit/roster either way, no email and no DM.