Why it matters
Why it matters
If you pay for more than one AI subscription, you can make them work as one team without a new tool.
My read: try it, if you pay for more than one AI.
Why I tried it
99%.
On a Saturday.
Claude hit 99% of my weekly limit on a Saturday. I was paying for Claude, ChatGPT, Cursor and Gemini anyway, so the work had three more places to go.
So I stopped asking my best model to do everything. It plans now. Everyone else builds.
How it runs
Rival companies.
One group chat.
Nobody asked them.
The setup, from the reel.
Me
Questions for me land on one page. I answer when I'm ready, all at once. Merges, deploys and anything that costs money still wait for my yes.
The boss
Claude Code / Opus 5.5
One session. It reads the plan, writes briefs and reads reports. It never writes a line of code.
Middle management
Claude Code / Sonnet 5.5
Two sessions, one per stream of work. Each takes a brief, starts the workers and sends a short report back up.
The workers
Each one is a single shell command, started when it is needed and gone when it is done. No harness running in between.
- Cursor's Composer
Composer 2.5
Builds, in its own worktree.
- A Claude Code subagent
Sonnet
Checks the work.
- Gemini, through Antigravity
Gemini 4 Argon
Reads my notes, read only.
- Codex
Gets its questions in an existing thread, the same one I see in my ChatGPT app.
Route it however you want.
Who builds is just a setting. I route by usage limits: when my ChatGPT limit drops too low, more work goes to Composer; when Cursor runs low, more goes back to Claude. That is why Composer is the main builder right now. Next week it might be someone else.
Next on my list.
Not built yetThe boss doing that routing itself, reading the usage meters and moving the work before a limit hits.
agent -p --model composer-2.5 "<brief>"agy -p --model gemini-4-argon "<brief>"codex exec resume <thread id> "<question>"Rival companies, one group chat. Nobody asked them.
The work behind the call
What happened.
- Past week
- The setup has run my Worth a Shot Club work for about a week: one planning session over two lanes since 13 Oct, Composer as the main builder since 19 Oct.
- 19 Oct
codex exec resumereached an existing Codex thread and answered in it, read only.- 19 Oct
- A Haiku subagent started its own subagent. So level 4 works too, if you enjoy that sort of thing.
- This reel
- This reel was made the same way: a Sonnet builder drew and rendered every version on a second machine while I gave notes in one window.
The catches
It has its
management issues.
Here’s what tripped me up and what I changed.
Name the model every time.
I asked agy for Argon and quietly got Gemini 3.8 Flash, its default. The answer header gave it away.
The fix --model on every call, and ask for the model name in the first line of the answer.
Resuming a thread can switch models.
A resumed Codex thread ran on a different model than the one it was made with.
The fix Same fix: name it.
Headless runs can't ask for permission.
My first headless agy runs produced nothing, because every tool needs a yes and nobody was there to give it.
The fix Allow the read-only tools explicitly. Never skip permissions, unless you want an agent with your credit card.
Too many reports.
In the first week the boss sent me way more than fifteen updates in ninety minutes, while the lanes asked me their own questions on top.
The fix Reports go to a status page, and only my decisions come to me, in one batch.
One writer per worktree.
Two Codex jobs in one worktree trip over each other.
The fix One writing job per worktree, start the next when the last one is really done.
A dead machine looks like a slow agent.
A background builder waited ten minutes on a render machine that refused every login, politely, without telling anyone.
The fix Check what a long job is actually doing, not just whether it is still running.
Try it.
Absolutely try this. If you pay for multiple subscriptions, combine their strengths in one workstream.
What this tells you
This is one person's setup, run for about a week. I did not measure time or cost against working without it, and your limits and tools will differ.
The setup keeps changing, so this page is dated.
Do it yourself
Give them
their assignments.
The guide from the reel. Start with the tools you already pay for.
Pick one Claude Code session as the boss.
Tell it in its CLAUDE.md that it plans, writes briefs and reads reports, and writes no code itself.
Give it lanes.
Sonnet subagents or extra sessions, one per stream of work, each taking a brief and reporting back in a few lines.
Install the command-line tools of what you already pay for
and log in once: Cursor's agent CLI, the Codex CLI, Antigravity's
agy.Let the lanes call them as one-liners
from the shell, with the model named.
agent -p --model composer-2.5 "<brief>" agy -p --model gemini-4-argon "<brief>" codex exec resume <thread id> "<question>"Make every worker say which model it is
in the first line of its answer.
Allow read-only tools for headless runs.
Never use a skip-permissions flag.
One writer per worktree.
Keep reports short and put the details in files.
Your questions go to one page, in one batch.
Tools & sources
What I used.
- Tool
Claude CodeLevel 1 and level 2: the boss on Opus 5.5 and the two lanes on Sonnet 5.5.
- Tool
Cursor CLILevel 3: Composer builds in its own worktree, started from one command.
- Tool
CodexLevel 3: answers questions inside an existing Codex thread.
- Tool
Antigravity CLI
Level 3: Gemini reads my notes, read only, through the agy command.
How this reel was made
- Repo
HyperFrames
The reel is an HTML composition, rendered on a second machine.
- Skill
design-engineering
The launch-video rules the builder followed for the floating windows.
Yes, that's my AI avatar in the video. Same opinions, better lighting.