- jaypeters.dev
- Posts
- Multi-Agents, Codexes, and Human Emulation
Deep into multi-agent management this week, as I ramp back up into proper modern software development.
Plenty of AI terms ahead.
Topics for this week:
Parallel Parallels
I’m not quite to the pro-levels of running 15 agents at a time like the creator of Claude Code, but I did ramp up in agentic development this week.
I had GPT-5.2 work on a number of tasks to help with Wave while also learning how to run more agents at the same time.
Cursor automatically handles “best-of-N” development for you where you have one prompt, and it spins up 5 copies of GPT to all attempt the same task.
After they’re done, you get to pick the best output.
But what I really wanted to do was work on different things at the same time.
Cursor doesn’t handle that automatically outright.
If you try to open the same project in 2 different Cursor windows, Cursor will automatically switch back to the currently open project.
So there’s no way to run 2 different prompts on the same project with Cursor.
Which is probably good, that could create some nasty situations of one AI not realizing another AI is changing things.
For true parallelization, either some sort of virtualization would be necessary, like Cursor running in the cloud and working on a different branch, which may be possible with the browser-based version of Cursor.
But the multi-Cursor window solution on one computer requires using git worktree.
Git lets you branch code and keep changing in separate instances, but a branch is part of a single project.
Worktree lets you copy the entire project and treat it as a separate project with separate branches.
With a new worktree, it’s possible to open a new Cursor window, and have another AI get to work on a completely different prompt, not the 5 agents working on the same project from before.
Taking this to the next level, it means it’s possible to have 2 Cursor windows doing “best-of-N” on 2 different prompts.
Potentially 10 instances of an AI working at the same time.
Parallel parallels.
And good bye monthly token usage.
Maybe not, but I would imagine running that many agents at a time will immediately hit at least a few dollars worth of API compute in a single session.
I’m still early in my multi-agent development days so I’ll be trying out what works best for me in terms of management and budget, but this is one potential path.
Multiple Cursors working on multiple prompts.
Multi-Agent, Multi-App
Another way of having multiple AI agents work in parallel would be to use different programs at the same time.
For Wave, I was limiting myself to Cursor for development, but I did end up setting up the Claude desktop app so I could try out their new Cowork feature.
It took a few attempts to get my subscription reactivated and connected with the desktop app, but once reinstated, I had Claude Opus digging through files on my computer and reorganizing Wave screenshots and other random images that have been piling up for over a year now.
All this was happening in-between sessions of waiting for GPT to finish working on the current development task in Cursor.
And I was impressed with the results.
Cowork neatly cleaned up the files on my desktop.
I didn’t get much further than the desktop cleanup with Cowork, but off to a good start with Claude.
At one point during this, I introduced a third AI app because I still have the ChatGPT app running.
So while Cursor was developing, and Claude was organizing, I ended up asking ChatGPT a separate development question, totaling the active agents and apps to 3.
This multi-agent workflow could really be pushed even further because so I’ve only done this with dedicated apps.
There’s multiple CLI tools and web-based tools that could be brought into the fold and increase the parallelization to new levels.
Really the task I’ve had in mind for this is Codex.
GPT-5.2-Codex is supposed to be good for long-running tasks, and OpenAI provides a pretty generous quota for using it in the browser, so Codex in the browser will probably serve as a perpetual code review and refactor machine, looking for obvious and easy fixes on big projects like Wave, and doing so all within the cost of the subscription, no extra API costs needed.
So this week, I started to get a taste of the multi-agent craziness we are about to enter as these AI models and tools get better.
We’re practically already there, the real bottleneck is coming up with the ideas and orchestrating it all.
More to come on multi-agent workflows, but for now Cursor, Claude, and ChatGPT is a very solid start.
Codex codex Codex
Digging through agentic recommendations on X, I see mostly Claude Code and Codex as the preferred tools.
Determining the context of “Codex’ is a little tricky right now, though.
There are basically 3 versions of Codex from OpenAI.
First, there’s the Codex CLI (command line interface) that you can download and run on your computer much like Claude Code.
Then there’s codex the AI model, like GPT-5.2-Codex.
Finally there’s the https://chatgpt.com/codex version of Codex which is basically an extension of Codex CLI except that it runs Codex in the cloud for you in it’s own virtual machine instance so that it can process in the background.
As you can probably see, the hard part here is that these are all interchangeable to some degree.
“Using codex” could refer to the locally running Codex CLI that uses an API, but that doesn’t mean it’s using the codex API model, just the Codex app.
It could be Codex using the normal GPT-5.2 model, not the codex model.
On the other hand, using Codex could refer to using the codex AI model in some other harness like Cursor, since that is also an option.
Or it could be both, using the Codex app to run the codex model.
One or more forms of Codex doing development somewhere.
On top of that, there’s the confusing distinction of using Codex at what cost?
Using AI in the browser is generally free but using AI through the API costs per API call.
However, OpenAI provides a quota included in a paid subscription for Codex usage that doesn’t use API costs.
So the Codex CLI can be setup to use an OpenAI subscription for Codex-based development cost, which is separate from the API costs of using the codex AI model.
I had been avoiding using Codex altogether because of all this.
I didn’t want to accidentally start racking up API costs because I let something run over night and it didn’t go as planned.
But I’m pretty sure I’ve gotten a grasp on it now.
Learning Codex CLI can be configured to use my subscription quota, means I’m not stuck with the browser to take advantage of that quota, and can worry less about token costs.
It’s much easier to kick off a scheduled script in a terminal than it is to automate things in browser, although browser automation is getting easier by the day…
As if there weren’t enough AI tools to use, Codex is another one to add to the list, but seeing it referenced so much online, I figured it was finally time to give it a shot, and after some digging this week, I feel comfortable doing just that.
I’m sure that was perfectly not confusing learning how to setup codex with Codex or Codex.
ClawdBot
When I was thinking I was limited to using Codex in the browser only, I was debating ways to automate browser interaction.
There’s a few options out there now, even OpenAI and Perplexity released AI-integrated browsers for just such use cases, though I think they are severely limiting Atlas’ and Comet’s abilities because they don’t want to make it even easier to create spam through these tools.
Anthropic’s Cowork feature comes with a browser extension so it can take control of the browser as well.
But a new tool popped up on my radar this week that takes the browser automation idea, and even the Cowork concept to a whole new level.
ClawdBot is an open source alternative to controlling everything on your computer through AI models like Claude Opus or GPT.
Once installed and connected, you can message your agent through Discord, Whatsapp, Telegram, Slack, etc., and it can do what you tell it to do on your computer without you having to be there, including operate the browser for you.
I’m not quite ready to give an AI full control of my computer yet, so this may be more of a VM situation to test it out, but I would be super excited to be able to delegate tasks on my computer through messages.
Better yet, could set up the automations I’ve been thinking about but haven’t taken the time to properly setup.
Just have the agent figure out how to navigate through the machine for me and do the task.
Really powerful stuff.
ClawdBot may be a good weekend project so more on that if I do get it up and running.
Human Emulation
I saw that an xAI engineer was on a podcast and a few days later announced he no longer worked at xAI, but I hadn’t seen the reason or if it was just coincidence.
Sounds like he revealed the future roadmap for xAI so he was asked to leave.
A “full human emulator” is the plan.
I’m not sure leaking this is really a shock, this seems to be the general plan for most companies building AI.
Create software that can act like a human.
It’s certainly an interest spin on the concept, though.
Human emulator is a new one for me anyway.
But it makes sense, it’s concept I’ve been thinking about for a while.
“MacroHard” is Elon’s plan for an entirely AI-operated software enterprise run by human emulators.
And each week it’s feeling like that could be closer to reality.
Grok 4.20 hasn’t released yet, but supposedly it won a stock trading competition against other AI models.
Rumors are the GPT-5.3 is possibly coming out next week and seems to be performing very well.
Even an update for Gemini may be right around the corner too.
Lots more options about to be added to the human emulation toolbox.
We’ll see what comes out next week.
And that’s it for this week. Lots of agent coordination and model management and potential human emulation.
If you want to start a newsletter like this on beehiiv and support me in the process, here’s my referral link: https://www.beehiiv.com/?via=jay-peters.
I also have a Free Skool community if you want to join for n8n workflows and more AI talk: https://www.skool.com/learn-automation-ai/about?ref=1c7ba6137dfc45878406f6f6fcf2c316
Let me know what else I missed! I’d love to hear your feedback at @jaypetersdotdev or email [email protected].
Thanks for reading!