← ClaudeAtlas

orchestrationlisted

Delegation policy for kelpie — when delegating actually pays, which role fits, and when to run a check instead of a verifier. Consult before spawning a subagent or writing a workflow script in a project where kelpie is installed.
seahsky/kelpie · ★ 1 · AI & Automation · score 69
Install: claude install-skill seahsky/kelpie
kelpie ships two roles as plugin agents (invoke as `kelpie:<role>`): `mech-executor` and `verifier`. It used to ship five. The other three were removed after kelpie's own benchmark measured them, and the measurements are in this file because they change what you should do, not as trivia. ## Default to not delegating Delegation is a cost, paid up front, recovered only in specific shapes. Two measurements from kelpie's Stage 2 run (260 scored trials): - Over 180 trials on ordinary coding tasks, Opus 5 spawned a subagent **zero times**, with kelpie installed and without. Installing the plugin changed nothing, because the model never reached for the mechanism. - On 40 trials where delegation was forced through a workflow, the delegated arm cost **6.37x** the plain prompt for the **same 100% pass rate**. The gap decomposes as 7.27x the tokens against a 1.14x cheaper blended price. Output tokens, at 18.81x volume, drove it — not re-reading context. The lesson is not "never delegate." It is that a cheap worker does not make a task cheaper if delegating adds turns. A spawn that runs a 9-turn internal loop to answer one question costs more than answering it inline, at any tier. **Delegate when the work does not fit one context, or when independence is the point:** - More files than one context holds, each needing the same treatment. - Wide read-only search where you want the conclusion, not the file dumps. - A check that must not be done by whoever wrote the thing. **Do n