GitHub - ptmrio/harness-subagent: Dispatch Claude Code, Codex, or Grok as a one-shot subagent, then synthesize. · GitHub
/" data-turbo-transient="true" />
Skip to content
Search/
Sign in<br>Sign upAppearance settings
You signed in with another tab or window. Reload to refresh your session.<br>You signed out in another tab or window. Reload to refresh your session.<br>You switched accounts on another tab or window. Reload to refresh your session.
Dismiss alert
{{ message }}
ptmrio
harness-subagent
Public
Notifications<br>You must be signed in to change notification settings
Fork
Star
main
BranchesTags
Go to file
CodeOpen more actions menu
Latest commit
History<br>4 Commits<br>4 Commits
Folders and files<br>NameNameLast commit message<br>Last commit date<br>assets
assets
evals
evals
references
references
scripts
scripts
.gitattributes
.gitattributes
.gitignore
.gitignore
LICENSE
LICENSE
README.md
README.md
SKILL.md
SKILL.md
View all files
Repository files navigation
harness-subagent
Orchestration for coding agents — stay in the parent, outsource to other harnesses, then synthesize.
Ever wanted to call Codex from Claude Code ?
Ever wanted Grok Build to invoke Claude Code ?
Ever wanted Claude Code to send a diff to GPT for a second opinion?
Ever ran out of Grok Bot usage after heavy tasks?
That's this skill. You have an orchestrator — Grok Bot , Cursor Agent / cursor-agent , Claude Code, Codex, whoever you are already in. It writes a brief, dispatches another coding-agent harness as a one-shot subagent , then the parent synthesizes.
The other harness is not an oracle. A model reviewing its own work reproduces its own blind spots; a differently-trained harness does not. That worth is destroyed the moment you forward its answer without judging it.
This is a Claude skill in the Agent Skills format. The same protocol works from Cursor Agent, cursor-agent , Grok Bot (if that Bot can run a CLI), and any other agent that can run a CLI in the background.
Why this exists
I built it as orchestration for Grok Bot .
Grok Bot is Cursor’s persistent-agent product — its own weekly usage meter on your Cursor account. That meter is not Cursor Agent’s Grok 4.6 pool, and it is not Grok Build CLI. It resets weekly. It is drawn down by agent steps and tokens, not by message count: one heavy task can burn a large share of the week, even on Ultra.
This skill is how I stop spending that meter on the heavy slice. While the Bot still has usage, it writes a short brief and outsources the long run to Claude Code, Codex, or Grok Build CLI (those CLIs bill their own accounts). The Bot synthesizes. That is conservation.
If Grok Bot weekly usage is already gone, the Bot cannot orchestrate. Wait for the weekly reset, use on-demand if you have it, or switch the parent to Cursor Agent / Claude Code / Codex.
cursor-agent is a different parent (Cursor’s CLI). Same protocol, separate product from Grok Bot.
Why I use it this way
These are my defaults, not the protocol. The skill will not pick a harness from the job type. You name the CLI, or you pin defaults in a user config file the parent is instructed to read, or it asks once.
I find Opus (Claude Code) superior at UI work — layout, interaction, the thing on the screen. I find GPT / Codex superior at reviewing , especially visual review : screenshots plus the named CSS/JS, then a verdict on whether the defect is real.
So my default loop is: ship the UI with Opus, capture shots, dispatch Codex to confirm or refute, then the parent decides. Not the other way around, and never a paste-through of the subagent's report.
Going the other direction is the same idea. If you are already in Codex or Grok Build and the job is a UI slice, dispatch Claude Code / Opus to implement it.
Recommended use
Author defaults (not the protocol):
Job<br>I dispatch
Heavy slice while I am still in Grok Bot (save Bot usage)<br>Claude Code, Codex, or Grok Build CLI — brief, then synthesize
UI implementation, layout, interaction<br>Claude Code (Opus)
Diff / correctness review<br>Codex
Visual review (screenshots + named sources)<br>Codex
Pressure-test a plan (assume it is flawed)<br>a different harness than the one that wrote it
Stuck bug, two fixes already failed<br>a different harness than the parent
How to ask (from whatever agent you are in):
Orchestrate this — outsource the slice to Codex.
From Grok Bot: have Claude Code do this UI slice — named paths only.
Ask Codex to review this diff.
Get a visual review from Codex of the screenshots I just took.
Ask Grok Build to try to refute this plan.
Ask Codex from Claude Code / have Grok Build invoke Claude Code.
Worth a run: second opinions, adversarial review of plans, visual confirmation, unstuck diagnosis, a bounded implement slice assigned to that harness.
Not worth a run: naming, style, formatting, or anything the parent can already answer from context.
The parent writes a bounded brief (named paths and a tight investigation — not a pasted...