Asking One Assistant to Direct Another, Tested
Codex, Grok Build and Hermes each asked to run Claude Code on a small job: what stopped them, what let them through, and what each handoff proved
Tested 17 November 2026 on one Mac and one second computer (Linux, through hc), from Claude Code, Codex, Grok Build and Hermes; Antigravity as the one directed. Claude Code 2.1.326; Codex CLI 0.166.0 (gpt-6.1-sol, low effort); Grok Build 1.0.70 (grok-4.6-build); Hermes Agent v0.23.2 on ChatGPT Terra (gpt-5.6-terra).
Some of the assistants in this series can start programs on your computer, and another assistant is a program. So you can ask one assistant to direct another: tell Codex, Grok or Hermes to “ask Claude to write this, run it, and tell me what happened.” This page reports what happened when that was tried on a deliberately small job, and where each attempt stopped.
It is a test of the handoff, not of the quality of the work. The job was trivial on purpose, so a failure could only be the handoff.
The job
A scratch folder holding one file, sample.csv:
item,qty,price
seed packets,12,2.50
trowel,2,14.00
gloves,3,6.75
The brief to each assistant (the full text is in the research folder):
Ask Claude Code, which is installed on this computer as the command
claude, to write a small Python programtotal.pythat readssample.csvand prints the grand total of qty times price. […] Then run the program yourself and report the exact command you used to start Claude, what Claude said, the output oftotal.py, and the files that changed. Do not commit or publish anything. If Claude Code needs sign-in or approval, or your own sandbox blocks it, stop and tell me exactly where it stopped.
The right answer is 78.25 (30.00 + 28.00 + 20.25). Every total.py was also run by hand afterwards.
What happened
| Assistant | First attempt | What got it through |
|---|---|---|
| Hermes Agent | Worked, about a minute. | Nothing extra. |
| Codex CLI | Stopped: Claude answered “Not logged in”. | Letting its sandbox reach the network and Claude’s settings folder; then Claude’s own edit permission. Third attempt worked. |
Claude Code, directing Antigravity through hc |
Stopped at Antigravity’s “trust this folder” dialog, which I left for the owner to answer. | The owner said to trust the scratch folder. Then it worked: 78.25, no further prompts (Antigravity’s auto mode). |
Claude Code on this Mac, directing Claude Code on a second computer (hc exec --on) |
Worked, about a minute. The remote Claude said it could not run the program (it needed approval) and worked out 78.25 by hand; I ran it on that computer and it printed 78.25. | Needed the full path to claude there, since it is not on that machine’s default path. |
| Grok Build | Stopped: “a headless run cannot ask for approval”. | Its own --permission-mode auto. Worked, about two minutes. |
Three different stops. Each is a different layer of “may this run?”:
- The assistant’s sandbox. Codex runs commands in a sandbox. Inside it, Claude could not find its sign-in and said “Not logged in”, though it was signed in outside. Opening the sandbox to the network and to Claude’s folder fixed it; the test did not separate which of the two did.
- Claude’s own permission. Claude Code asks before writing a file. Run once with
-p(print and exit), it cannot ask, so it refused and said so. The brief then told Codex to start it with--permission-mode acceptEdits, which allows edits and nothing broader. - The coordinator’s approval. Grok Build, run once from a script, would not run the shell commands because nobody was there to approve them. Rules to allow named commands did not help here, because Grok chained several commands together and a rule written for one did not match the chain. Its
automode, which Grok decides for itself, did.
All three reported their stop accurately. Codex quoted “Not logged in” and said no file was created; Grok said a command “was not run”; Claude said it could not write. None claimed success or invented a result. That is the main thing to ask of a coordinator: tell me where you stopped.
None of the stops was removed by a risky switch. The flags that skip every check (Codex’s bypass flag, Grok’s --always-approve, Hermes’s --yolo) were not used. What worked was widening one thing at a time.
Continuing a Claude session, or starting a new one
Hermes was asked to put the same question to Claude twice in the same folder: “Which file did you write earlier, and what total does it print?”
- With
--continue(the most recent session in that folder), Claude answered: it wrotetotal.pyand it prints 78.25. - As a new session, Claude said it had no record of writing anything, but found
total.pyin the folder and described what it did.
Both give a useful answer on a small folder. The difference is memory: a continued session knows what it decided and why; a new one only sees what is left in the files. When you want the earlier reasoning, say “continue”.
What this does and does not show
- One run per route. This shows the handoff can work, not how often.
- The route was given. Each brief named the command and the flag to use. It does not show that an assistant would find the way on its own.
- Command-line tools only. Grok Bot and ChatGPT’s computer use and connected computers were not run; they need account access this test did not have. The vendor documents describe them (for example, Grok Bot works from a cloud computer that can also run commands on a desktop you enable; ChatGPT can control another computer you pair it with), but their documentation shows what exists, not that this handoff succeeds.
- Claude as coordinator: one trial, Claude Code sending the job to a fresh Antigravity (Gemini 3.8 Flash) in its own terminal with the
hccommand. I did not touch an Antigravity session that was already in use. - Not tested: the host going offline, and a marker file to tell the local machine from a cloud one.
Habits that follow
- Name the route. Say which command or interface, and the folder.
- Say what to do when blocked. “Stop and tell me exactly where” got an honest report every time.
- Widen one permission at a time, in a scratch folder first, and prefer an assistant’s own middle setting over the switch that skips every check.
- Check the result yourself. Ask for the file and the command output, then run it.
- Say continue or new. They are different tests and different memories.
Research, saved runs and scripts: the 2026-11-17-computer-control folder (a public export of the repository).