Claude Code + Codex CLI + Gemini CLI = ??

So I recently found a tool that may fix many  of the problems that you've been having with   some of your AI coding agents. But  to give some background to this,   over the last few weeks there has  been a trend happening where many   people have been combining different  coding agents CLI tools together. So,   for example, combining Claude Code and Codex  CLI, Claude Code and Gemini CLI and so forth. I actually made a previous video about this  called Codex CLI Just Fixed Claude Code over here,   where I basically used Codex CLI to critique  the implementation that Claude Code came up   with and then pass that information back to Claude  Code to then fix its own implementation and then   got it to critique again until I was happy with  the final result. But in this particular case,   I was a human in the loop, copying  information back and forth. And some   people have automated this process before  where they basically got Claude Code to   use Codex CLI as an MCP server, as  described over here in this post. And one month ago, this person wrote a  post where they basically combined Claude   Code with Gemini CLI to achieve a better  result than using Claude Code individually.

Because they basically said that Claude Code  excels in detailed conversational coding,   but it hits its context limits with large  projects. Gemini CLI, with its massive context   window and planning capabilities, can digest  entire codebases and provide strategic insights. And they basically gave the example that they  were using a Flutter project and users reported   that switching between tabs would cause a crash  in the application. And then this actually felt   like a system-wide issue that would sometimes  happen on a profile tab, sometimes on Settings,   but always after rapid tab switching. So they  basically got Claude Code to then call Gemini CLI. And then Gemini CLI identified that some  controllers were initialized in different ways   on different routes over here. And then Claude  Code worked to implement a solution based on the   insight that Gemini CLI uncovered. And then  after Claude Code finished all its changes,   they asked Gemini CLI to review those  changes as an external peer and checked that. The fix was consistent across the entire codebase  because of its massive context window. And this

is kind of the thing that I've been doing  with my application HyperWhisper over here,   there's a coupon code down below if you're  interested in buying it. And basically what   would happen is I would like get Claude Code  to do a fix or like add some kind of thing. It wouldn't work after testing it, then I'd  give it to Codex CLI, it still wouldn't work.    And then I would give it to Gemini CLI as  a last result. And more often than not,   somewhere along the chain, one of the  coding agents would actually fix it.

But this project over here,  which is a fork of OpenAI Codex,   kind of implements that approach for  you. So basically, if you scroll down,   you can see it has a couple of features over  here, like browser integration, Diff Viewer.    But the most important thing that I want to  focus on is the multi-agent commands over here. So basically what you can do, and I'm  going to show you an example soon,   is you can have multiple agents running  in parallel, Claude Code, Gemini CLI,   and Codex CLI, all in parallel on the same  codebase trying to do the same thing. So for   example, all of them can come up with a plan  for something and then code will combine all   those plans together in a big unified plan. Or  you can have them all try and solve an issue   in parallel and basically the agent that gets  it solved the fastest will end up having its   solution like as the preferred solution because  of this paper that they referenced over here. Or you can have all of them write code in parallel  and then it creates multiple git worktrees and   then implements the most optimal solution.   And if you don't know what git worktrees are,   because they're kind of a niche feature,  they basically allow you to have two   branches checked out at the same time. And  you can read through the rest of this to   understand how it works now before going  through an example of how you can use it. If you do want to follow along with  the latest AI coding news, tools,   and best practices and learn  some viral marketing as well,   then I cover a lot of that in my Skool  community over here. There should be a   link down below if you want to join it. And if  you do join it, then you get a 100% discount. Basically free access to all the applications  and tools that I personally make for being   part of the community and you get to  engage with a lot of discussions about   these kind of things and learn from others  as well. Now, basically, to get started,

you can either run this command over here  and then it will immediately launch code,   or you can actually install it on your machine  and then use the command code or coder instead.    And now you should see something like this and  say, yes, allow Codex to work in this folder. And this is because it's basically  a fork of Codex. Press Enter and   you will see something like this. Now  you may see a horrible light theme. So you can do /theme, press enter and then  switch between any of the themes over here.    So I'm going to go for this Carbon Night theme.   Now you can see currently the model is GPT-5,   Reasoning: medium and I will change that to  high because like GPT-5 is the main model   being used since it's a fork of Codex, so it  will work in the same way that Codex works on   your machine. So I can do /reasoning high,  and now the reasoning effort is high. And

then basically what I'm going to do is I'm going  to plan a fix or change, so I can write /plan. And basically I'm going to use my application  Tensor AI over here. So this is an AI news   application. You can download it  for free using the link down below. And basically I got a bug report that  came in a few days ago where someone   said there's an issue with swiping over  here and there's another issue. So I can   just copy over this bug report and then say  solve this and then paste this in over here,   press Enter and now you can see that it's going  to spawn three different agents in parallel.    Gemini CLI that you should have installed on  your machine, Codex CLI and also Claude Code. So you can see it has its own plan over here.   And now it's launching the planning agents.    And if you scroll up over here, then you  can see that using an agent tool to start   a batch of agents running in parallel,  which is Claude, Gemini and Codex CLI. So this may take a little while because  I do have it on Reasoning effort high,   but I have done this previously  in a different codebase before.    And if I scroll back up, you can see that  it started the three agents over here and   then it waited some time. And now this  is the response of one of the agents. So if I scroll down then there should  be a file that it's been saved to,   so I can click on this file and then see  this particular response that one of the   agents gave. And then I can also see the  other planning responses over here for this   feature of adding system audio recording.   And basically if I then go back to code,   I can scroll down and it says that  all three agents have been completed. And I can actually make sure that  it called Claude, for example,   because I can go to the same folder, run  Claude over here. And then if I do /resume,   then I can see that there has been  a previous conversation over here. And this is basically the conversation that  code did itself. So you can see these are   instructions that Code gave to Claude Code and  then it basically did the execution. And then   you can see that this output of Claude Code was  saved into this file over here, the result file. Anyways, you can see the other agents  that we have running for the other plan,   Gemini has been completed.   Code has finished running.    And that's referring to like Codex CLI,  because this is a fork of Codex CLI. And now GPT-5 high, which I have running,   will combine all these plans  together into one big final plan. So you can see all three planning agents have  been completed. Here's a synthesized plan that   merges their strongest recommendations all over  here. The plan is pretty comprehensive, so you   can get it to ignore some things like this over  here, or things that you don't think are relevant.

So basically we'll try the other two commands now.    So I'm going to run solve over here.   So do /solve "implement phase one". And now what it's going to do is run all  three agents in parallel. And whichever   agent ends up finishing first, it  will use that particular solution.    And now you can see it's fired off all  three agents in parallel over here. And it has like a nice little loading animation.   You can see it includes the previous agents that   ran before. Currently Claude is running,  Codex is running and Gemini CLI is running. Not exactly sure what this means,   but like it looks kind of cool. And now  whilst we're waiting for this to complete,

I can go back to the other project and then  basically use the other command over here,   which is actually code. And this will like  basically use a consensus between Claude, Gemini   and GPT-5 and it will do across multiple git  worktrees and implement the most optimal solution. So I can do like I can do /code "implement  the plan that you described above",   press Enter. And now I should see all  three agents fly off in parallel again,   but on different worktrees. And now you can see  that all three agents are now running in parallel. And I should be able to see the different  git worktrees. I can do git worktree list.    And I can see that there are three worktrees  over here, Claude implementation over here,   Codex implementation over here, and then  the Gemini implementation over here.

Now basically, whilst we're waiting for both of  these to run, what I can do is I can show you   why I think this works well. Basically, you can  imagine your codebase as being like dozens or even   hundreds of files all scattered over here. And you  want a coding agent to make one particular change. So you ask Claude Code to make a change  for you, and it kind of makes a change,   like in this region over here. So it's  managed to capture all the files related   to that particular feature that you wanted.   But sometimes what happens is because of   the way it's indexing the codebase or gathering  files or tool calling to navigate the codebase,   it ends up missing one of the files just  on the edge, for example, over here. And that particular file could have been pretty  relevant or pretty important for the solution. So   what ends up happening is that Claude Code often  writes a new function from scratch, even though   you had that particular function in the codebase  already. Now, when you run multiple agents in   parallel, so you can then run, for example, Codex  CLI in parallel on the same particular plan or   prompt, you kind of have a different overlap  over here, so maybe it looks more like this,   and it captures a lot of the relevant files for  that particular implementation that Claude Code   did not manage to capture because it navigates  the codebase in a different way, calls tools   and so forth, because it  is fundamentally trained in a different way. And then you have Gemini CLI that does a much  larger grab of the codebase because it has a   much larger context window, captures many more  files, and then is able to consider more things.    And it's basically captured a lot of files that  may not be relevant as well, but also captured   files that may be relevant too. So this is what  I like to imagine that's happening behind the   scenes when you're implementing a particular  plan with all three coding agents in parallel.

And if we go back to our terminal, you can see the  implementation /code is still running over here,   but /solve has actually been completed over  here. So I can press CTRL D to see the diff   of what's changed over here. I can press E to  then, like, actually have it explained to me. Press left and right to, like, navigate between  the tabs of basically the changes that have been   made over here. And then I can also undo any  changes by pressing U. If I press CTRL D again   to remove the diff, I can see this explanation  appearing of the E command that I pressed earlier. Then you can see that basically in the solve mode,   what happened is it launched all three agents  in parallel over here and then said all three   agents are completed. I applied the first  agent's concrete recommendations immediately,   because I think that was the one to finish  first, and then cross-checked with the other   two agents to ensure coverage and catch gaps. And  that's basically what the solve does over here. It runs all three agents, but whichever  agent like had the fastest solution is   preferred because of this particular  paper that they reference over here   that you can read through. And now as  for the actual code over here where

it's running on different worktrees, we  can see that Gemini is done over here.    Codex code is still running because code is  a fork of Codex, which is worth remembering. And Claude Code is still running over here.   Okay, so now I think the agents are done   over here because it's saying comparing agent  results and it's running this built-in command   to basically compare the implementations across  the different git worktrees. And I think that   GPT-5 reasoning mode medium as specified over here  is now making a comparison of all the results. And you can see it's writing this out  over here and it's describing, hey,   these are the different worktrees, this  is what each agent implemented. And then   it comes up with a comparison, it talks  about the strengths of each approach and   extra touches that certain agents have  implemented and then some new components   over here and so forth. And now it's saying  what I can do next is merge and integrate. So option A, which is the fastest, would be  to implement the synthesized plan already in   a new branch and commit in small steps, open  a diff for your review. Option B is reuse. If your Codex agent finishes with commits,  I'll inspect its worktree and cherry-pick the   best parts, blah blah into the feature  branch, blending with Gemini and Claude's   better API/UX decisions. And I think that's  true because Claude usually makes better UX   decisions than Codex CLI does. Or I can do  option C which is basically rerun them all. And I think the reason it gave this particular  approach or this particular options is that   code is actually still running for some  reason. I guess that's because this was   like a humongous task that I gave it to  do. So I should probably increase the   timeout in some way for code to finish  running and then do the comparison. So I'm going to say go for option A presenter.   So you can see it made two commits over here,   it made an audio commit and a chore commit.    And then basically it just gives me  all this other information as well.

So now I can run Xcode and then actually run the  application over here. Turn on the System audio   option. I wish the option did appear like when  you actually explicitly enable that mode perhaps. But anyway, we can then quit and reopen  the application. But I should have been   more specific in my instructions by saying  that when they're in modes for example and if   they choose system audio, then it should  only use the system audio. But anyway,   now let's check whether this actually works and  it's recording my system audio as I requested. So let me open up YouTube and then  play a video. So like this video   looks pretty interesting and let me  skip to some part of the video. It   is in Japanese so I will like make  sure this language is on automatic. And now if I press option space,  stop playing the video over here.    So you could see it was actually  reacting to the computer audio and   it's actually transcribed what was said on  the computer. So try to paste it somewhere. If I go to History then I can see right over  here and I can see it actually recorded what   was on my computer. So I think this is really  good. Like it basically implemented it perfectly. And I think that's because it basically combined  many different approaches together by seeing what   all the coding agents came up with. So I will be  trying out more over the coming days and sharing

my thoughts in my community that you can join  using the link in the description down below. But basically my initial thoughts are right now  that if you're starting a brand new project,   you may want to use this for planning.   But like you don't want to use it for   the initial implementation. I think you  can just use Claude Code or Codex CLI   or something instead for the initial  implementation of a brand new project. But when your project gets bigger and bigger  and you have like dozens or even hundreds of   files all doing different things and you  want to implement like a specific feature,   it's often just better to like set up a couple  agents in the background using something like   code and then compare all the solutions.   Because nowadays, especially since AI coding   agents are getting quite good, each one has  its own strengths and weaknesses, of course,   and it's good to combine the strengths of  all of them together in some way. I don't   think for a while there will be one particular  good coding agent that excels at everything. I notice in cases like Claude Code is better  at design, Codex CLI is better at making macOS   applications, and Gemini CLI is better at  understanding big codebases. For example,   since for a while there won't be like one  universal coding agent that excels at everything,   I think it's best to combine approaches for  many different coding agents for tasks that   require it. If you're doing a very simple  task that isn't particularly complicated,   then you can just run Claude Code or Codex CLI. But I think if you're doing a really big  implementation of a particular feature,   then it is best to use multiple coding agents  and have them critique each other's work and   combine the best solutions of all of them.   One thing that is worth bearing in mind is   that since this project is a fork of Codex,  it will be running GPT-5 as the main agent   that reviews the results of all the other  agents: Gemini CLI, Codex, and also Claude   Code. So there may be some kind of like rate  limiting step or some bottleneck where you're   kind of limited by the reasoning capabilities of  GPT-5, but I have found them to be quite good. And if you do enjoy this kind of stuff and  following along with the latest AI coding tools,   news and strategy as well, do subscribe to the  channel because it lets me know that I should   be making more of these videos, and it also lets  the algorithm know that this is like good content.