Gemini CLI with Taylor Mullen | The Agent Factory Podcast

This week on the agent factory, >> it's the agentic loop in action.

>> I am not very good at like sitting down and reading.

>> My read later folder has its own gravitational pull >> with AI using AI these days. It is so easy to 10x yourself. 100x it.

That's the hard part.

>> Yeah. And my mom watches these.

So, mom, I'm still alive.

Hi everyone and welcome to the Agent Factory, the podcast that goes beyond the hype and dives into building production ready AI agents.

I'm Molly Pettit >> and I'm EMTT Mirage.

>> And today we're going to be diving deep into the Gemini CLI. So we used the CLI a bit in our last episode, including this live vibe coding demo that was a ton of fun. I suggest you check it out if you haven't seen it yet.

Um, I'm really excited to show it off in more detail.

And for anyone who hasn't heard or used a haven't hasn't heard of or used the Gemini CLI yet, it's a really powerful AI agent lives right in your command line.

And it's designed to help you with your everyday workflows.

>> Yeah. And I'm really excited because in this episode, we're actually going to show you how to use Gemini CLI for getting up to speed on a new codebase, supercharging your research and how you can use it to integrate it into your own automations.

>> Yeah. And uh to help us do that, we're very excited to have Taylor Mullen, the creator of Gemini CLI with us today.

We're going to pick his brain about the philosophy behind the tool and what's next on the road map, which I'm very excited to hear about.

>> Yeah, I can't wait. I mean, I've been using the CLI now for a few weeks and I already feel like it's become a mainstay in my workflow. What about you, Molly?

What's your take?

>> Oh, yeah. No, I'm totally sold.

Um, for me, the biggest thing is that it lives like in the terminal. It's where I do my work.

Uh, and I was also able to add it directly to VS Code, the IDE I use.

Uh, so that was super handy.

>> Yeah, I feel like having it integrated into so many different tools makes it indispensable.

And also on that note, having Taylor with us here today who's going to be able to tell us even more about the background and philosophy philosophy behind Gemini CLI will be invaluable.

But before we talk to Taylor, let's get hands-on with some demos on the factory floor.

Let's tackle a universal developer problem.

Um, which is onboarding onto a new codebase.

I think all of us have had to do this at some point. Uh, I'm going to go ahead and show how the Gemini CLI can help speed up this onboarding process.

>> Okay. Yeah. So, codebase exploration, that sounds pretty cool. And I feel like this will also be a really good showcase for Gemini CLI's massive 1 million token context window.

I feel like that's kind of the difference between asking questions about a single file versus maybe the entire architecture all at once.

>> Yeah. Yeah, totally. Um, and so to start, I'm actually not even going to clone the repo myself. Um, I'm just going to ask the agent to do it for me.

So, I've decided to use Google's ADK repo uh for this example. So, ADK stands for agent development kit. So, I'm just going to say clone the Google ADK Python repo from GitHub and I'm gonna see what it does.

Okay. So, it's searching the web trying to find the right link.

Awesome. It found the right right link.

It's asking me for permission to clone it.

Go ahead.

Cool. Yeah. And this looks like it's the agentic loop in action. It's kind of reasoning that it needs a piece of information.

It then chooses a tool to get that information, it looks like, and then it uses that information to complete the original request.

And then also at the same time it asks you if you want to proceed checking back in.

So >> yeah, this is a super helpful workflow.

I think the way I want to tackle this is let's imagine that we're like a new contributor to this project.

Um what do you think is the fir like what's the first thing you'd want to do if you were a new contributor to a project and you wanted to figure out how to go forward?

>> Feel like I would want to get a high level understanding of the project.

I think I learned best maybe from like a top- down perspective. Things like the tech stack, how code is structured, stuff like that.

>> Yeah. Yeah. Totally. Like in the past, in in the notsodistant past actually, uh you know, you might start with like reading a readme file, maybe contributing file.

Um you know, start piecing things together that way.

Uh and like I am not very good at like sitting down and reading a huge amount of documentation to be quite honest.

Um, so instead I'm going to ask the Gemini CLI to perform a highle audit of the entire directory.

So let's see.

I'm going to ask the Gemini CLI agent for a complete project overview.

I want it to tell me the purpose, the tech stack, and to analyze the architecture all in one go.

>> Yeah. And I feel like that's much more powerful than just relying on one readme document.

Like this way you can kind of point at the whole directory and it's cool because you're asking it to like synthesize everything the source build file the docs all of the code into a single coherent sort of overview right.

>> Yeah. Yeah. Excellent. Yeah.

It's like it's ingesting the entire directory to build out this like complete mental model of the project which is super helpful.

Um okay great.

starting to output.

This looks really good and helpful.

Yeah, it identified the purpose, the text stack, core dependencies, uh all that good stuff that's like helpful to kind of know at the outset.

>> Yeah, it's really verbose. What a what a good summary.

I wouldn't have been able to write it better myself, right?

Um and I guess like perhaps it'd be a bit helpful to save it somewhere persistent.

You know, I know you were looking into like MCP servers to help with this.

Did you find anything?

>> I did actually. Um, so I set up a Google Docs MCP for this. This allows me to access my Google Docs and uh create new files, create new folders, make changes, etc. Like from the Geminis CLI.

Yeah, I already actually created a ADK um folder inside of a projects folder.

So I can go ahead and ask the CLI to create a new file. Okay, so now I'm telling the agent to save that summary to my Google Drive. I'm giving it the specific folder and the file name I want it to use.

>> Kind of worth explaining what MCP servers are.

So, you know, for those of us who don't know, MCP stands for model context protocol.

It's essentially a way and a standard that lets you teach Gemini CLI or any agent for that matter kind of new tricks. It's like a plug-in system for the agent, right?

So, by running that Google Docs MCP server that Molly just mentioned, we've taught the agent a new set of tools for interacting with Google Drive. And it doesn't need to just be Google Drive or Google Docs.

You can, you know, plug into any MCP server from Gemini CLI. That's one of the cool powers of it. So, it has access to so many different toolkits.

>> Yeah. Exactly. Let's see. So, it is looking into this. It's asking if it can create a document. Yes.

Okay. It says it successfully created a document.

Ah, there it is.

This is great. So I think if we're going to contribute to the codebase like at some point um it would be worth understanding the project's culture and workflow right >> yeah I think every project seems to have a different way of contributing you know the maintainers like things to be done a certain way commits to be written a certain way and so obviously these are very important piece of information to kind of know as as a developer you would want to know this going into a codebase so I think it's a really good point so you don't want your first pull request to be rejected because you didn't follow like their commit message style, right?

>> Yeah.

And then so if I'm already going to be looking at like get commits and things, I think this is also a good opportunity to find out what the team has like recently been working on.

So I am going to ask it about both those things.

For this prompt, I'm asking the agent to analyze the last month of Git history.

I wanted to tell me what the team has been working on and how they work together, specifically their contribution style.

Okay. So, I'm going to go ahead and with this.

>> Yeah. And that's a perfect example of its kind of ability to do multilayer analysis.

It's not just running kind of git log as you would do maybe as a developer and then trying to parse through it one by one.

It's actually ingesting the entire output for the last month into its context window and then reasoning over it in two ways.

First, you know, to find the technical themes in the work uh and then second to infer the human process or the contribution culture.

So, I feel like this is really helpful because obviously we know the capabilities of LLM to like summarize and understand what's in the content.

So, taking all of that throwing it into the context window to give it enough information for contribution is obviously very helpful.

>> Okay. So let's let's do one more thing.

Um so I want to use this massive context window that Gem NLI provides us to do something that would be pretty timeconuming for a human which is a proactive uh code health check. I like I want to get a better sense of the top improvement opportunities in the codebase.

And for this final prompt, I'm asking the Gemini CLI to do a full code health check on the entire directory and recommend the single best first task for a new contributor based on that analysis.

All right, let's see what it comes up with here.

Okay, so it looks like the code base contains to-do a bunch of to-do comments which was helpful for figuring out kind of some next steps and then it gives me uh some suggestions as a new contributor and some reasons why.

Um, awesome. This is uh this is super helpful.

it would make it very easy for me to kind of just like jump in with something and then um I could also use the Jevi CLI to help me with those tasks which which is pretty cool.

So let's pause here and just do like a quick recap of what we did without leaving the terminal.

We had the agent clone a repository uh and then generate a complete personalized onboarding guide.

Uh, and now that I have this prioritized list of actionable tasks complete with um like a suggestion for maybe the first task task that would be good for me to do, I could start working on my first contribution.

All right, for our next demonstration, uh, we're going to shift gears from analyzing code to analyzing research.

admit I know this is a topic close to your heart as there's this absolute flood of new AI papers coming out every single day and it's tough to keep up with. Uh yeah, tell us about it.

>> This past week I decided to try to tackle some of that problem and build a solution using Gemini CLI while I was playing around with it. So I wanted to see if I could create like kind of an automated research assistant for myself.

>> Oh yeah, I love that. Um how'd you get started on it?

>> Yeah. Yeah. I think like uh just like with code you start with like a simple idea.

So I wanted to adopt a similar approach.

I had this directory full of research PDFs.

And so my first thought was you know I'll just get Gemini to explain them to me. So uh I tried a really simple prompt go through all the research papers in this directory and create a web page interactive explainer for each one. So I wanted to create this web page for every single one of my research papers uh and then create some kind of interactivity behind it.

So, I'm more visual learner.

>> I I bet that you'll need to iterate on this.

It's It's a good place to start, though.

>> Yeah, exactly. I mean, iteration's key.

So, here's a research paper that I've got, and then I booted up Gemini CLI, and it gave me a nice little greeting.

Uh, pasted in the prompt, and it worked, which was cool, obviously, but the outputs I felt like were a bit shallow.

Um, the interactivity was a bit minimal, and it kind of glossed over some of the deep technical details, which I wasn't super happy about. Um, and that's often most of the important pieces of uh, research paper, right? So, uh, it was a good start, but I knew I could do much much better.

>> Yeah, for sure. I mean, this is this is where like prompt engineering and iterating with the LLM comes into play.

>> Yeah. So, you know, I sat down, I went back to the drawing board and wrote a detailed multi-part prompt that acts as a full specification. I won't read the whole thing, but I gave it a new prime directive.

I essentially told that your highest priority is to represent all of the technical details from the paper with precision.

Try to avoid vagueness and simplification and instead focus on clarification.

If you actually want this prompt, uh I'll link it in the show notes and you can grab it for yourself.

>> Nice.

>> And then I got super specific about the output.

I realized asking for complex interactivity was a bit too vague and can lead to potentially messy code and or explanation.

Mhm.

>> So I instructed it to use static but clearly label diagrams, render math equations with a specific library and then only to use simple CSS for animations.

>> Nice. Yeah. Did it result in a better output and like how are you outputting this?

>> I wanted to build it in such a way where every research paper had its own individual self-contained HTML file.

And so as we begin to run with this, Gemini CLI approaches it in the same way it did for the single paper. But due to the more complex prompt, the output becomes a little bit more creative.

>> Cool. Yeah, that's very cool.

Um, it even has these like interactive examples it built using the libraries you specified in the prompt.

>> And I think you could even fine-tune the prompt a bit more to focus on like different technical aspects of papers.

Or if you just want like a more presentational view of it, perhaps you want to present your paper at a conference or something, you could perhaps use this, fine-tune the prompt, and then get your paper into a state where it's a little bit more presentation ready.

>> Yeah. And like I know that you well, first of all, research papers are going to continue to come out and I know you like to keep up on this stuff, so you're going to be running this again and digging up this prompt and kind of redoing this every time um might be a bit annoying.

So, here's the best part of this is I'm not actually going to have to copy and paste this huge prompt every time.

I can use the custom slashcomand Gemini CLI feature to save it.

And so, um, I can then, you know, run it anytime I need to.

>> Awesome. Yeah, I have not actually experimented much with these yet, but I'm excited to. I know you've created a a bunch though, right? Like, how do you go about that?

>> Yeah, super super easy. A lot easier than you probably think. So you kind of just jump into this Gemini folder which is created inside of uh your home directory.

Uh and then inside of there there's a folder called commands and then you can create your custom commands inside of here. Every custom command has its own TOML file. So this case here I'll just create a research_digesttoml file and the file is super simple.

It's just the name of the file, a title and then the prompt itself. So I created a simple title and copied the prompt that we used earlier and then I used the name for slashresarch digest. Uh, so now when I download a new paper, all I have to do is plop it into my folder, open the terminal for it, and then type that one thing forward/research digest.

>> Yeah, that's that's super slick.

Uh, you can just keep adding papers to it, and then just run that command the next time.

So easy.

>> Yeah, exactly. And and that's it.

I mean, a minute or so later, I have a brand new, beautifully formatted HTML file that serves as kind of like this rich interactive companion to the original paper.

probably making me a little bit more lazy when it comes to reading research papers, but overall accomplishing my goal in ingesting more information.

And it has all of the information that I want.

The full methodology, the equations are rendered perfectly and it's all incredibly easy to navigate.

>> Yeah, that's so that's so great.

And then like like so if you wanted to get really fancy, you could also automate the first part of the process too, right?

>> Yeah, it's a good point.

Yeah, absolutely. You could hook this into something like the archive MCP server which is available on GitHub uh and it automatically like downloads new papers in the fields that you're interested in.

The nice thing about Gemini CLI is that it has support for MCP servers.

So you could have fully automated pipelines that find you know the research for you.

It processes it and then it presents you with these kind of beautiful digests like every morning for example.

>> Amazing. Thanks so much for sharing it.

But now it's time for the agent industry polls where we chat about some of the recent developments in the world of agents.

>> So Lang Chain just dropped the 1.0 alpha.

They're refocusing the entire library around a new unified agent abstraction built on Langraph, which means production grade features like state management and human in the loop are now front and center.

>> Yeah. Yeah, and on that note, for anyone building rag based agents or just using embeddings in general, Google just released embedding Gemma, it's essentially this family of like open lightweight embedding models you can just run for yourself and essentially gives developers the ability to build ondevice flexible but privacy ccentric applications by generating embeddings of documents directly on the devices hardware.

So, and it it what's really cool is it also uses the same tokenizer as Gemma 3N for text processing.

So instead of relying on like a service, you get more control over the core of your agents retrieval system if that's what you're building. On that topic, there's also a new book out called agentic design patterns for building AI applications.

So it's one of the first real attempts to create a repository of education surrounding agent patterns, which is super cool. So this is particularly helpful in getting us to a production ready space and being comfortable with working with AI agents and deploying them to scale up.

>> Right. And not every agent task needs a giant model, right? Google also released Gemma 3270M, which is a tiny 270 million parameter model.

For builders, this is great for creating small, efficient sub aents that can handle simple tasks like data extraction or intent routing, making your overall system cheaper and faster.

We're also seeing more AI tools get integrated directly where we work.

The Gemini CLI is now built into the Zed code editor.

This means you can use it to explain code or generate snippet right in the editor without having to switch contexts similar to what I showed earlier in uh VS Code. And then Ait, I know you've been looking into helpful like agent resources. What are you seeing happening in the open source community?

>> Yeah, I'm so glad you asked.

So, a couple of good ones I wanted to share.

So, first I found this GitHub repository called 500 AI agents projects.

Uh, and it's essentially just like a huge categorized list of open- source agent projects that anyone can take a look at.

It's really useful if you're looking, you know, one for inspiration or just to see how other people are building agents and using different frameworks.

Then secondly, a team from Stanford put their LLM cheat sheet on GitHub. So I know for me like learning about LLMs and natural language processing uh has been you know quite challenging at times. So this is a really good starting point or a refresher on the fundamentals.

It covers things like attention and fine-tuning and pre-training and it also really presents it in a really simple nice visual uh easy to look at way and you can download this PDF, you know, read it on the train or wherever you are.

>> Yeah, very nice. I'm going to have to check that out. Um so yeah, there's a ton of great resources and updates there.

We'll make sure to link all of them in the show notes so you can explore them yourself.

>> Yeah. And now for the part of the show that we've both been waiting for.

We're going to shift gears and get a deep dive into the philosophy and future of Gemini CLI.

>> We'd now like to welcome to the podcast Taylor Mullen, creator of Gemini CLI.

Taylor, thank you so much for joining us.

>> Thanks for having me. This is fun to be here.

>> Yeah. So, let's let's start at the very beginning.

Um, could you tell us the origin story of the Gemini CLI?

like what was the initial spark or inspiration that led to it and how did things go from there? I'm very curious to hear.

>> Oh yeah, it's it's kind of crazy.

It started about a year and a half ago, weirdly enough, right? So long ago.

>> Wow.

>> But like at the time I was experimenting with multi- aent systems and trying to get like the most out of our LMS. And so at the time like I had the system where there's multiple personas.

They were talking to each other and they were doing a lot of cool things. And it was surfaced in a lot of different ways.

It was surfaced in the IDE. It was surfaced in a CLI. Um it was surfaced in the web.

And as I was playing with these things, the thing that really stuck for me was the CLI.

Um it just beca it was so easy to use.

It was so lightweight. Um but at the time like it took 30 requests like like one user request resulted in like 30 LM requests behind the scenes.

And those 30 requests was a lot was a lot for the time. It took a minute and a half to respond to anything. Um and then while the quality was there like this was something at the time that was just too much.

Like we also have limited context window.

Um and so in that realm um it being a CLI and it having all these constraints uh we scrapped the project and that was kind of challenging at the time. It was almost a little bit too early.

>> Yeah. Yeah, it was like it's just one of those things like, okay, no one's going to want to use this because it costs too much.

It couldn't absorb enough information and it took too long to respond.

Um, which is kind of a trippy thing like fast forwarding to today, right?

Like >> we're very much used to you have deep research like you click a button and you go off and you get a coffee and you come back and hopefully it's done.

And like it's it's crazy. It's so different.

Um but during that time like it was too early and then today like me Google and me um really trying to like build this future of developer tools like there was this ecosystem of CLIs that was just taking hold in the community right and it really proved that people are willing to wait people are willing to um use a CLI and they find it they found it as compelling as say I did back then it's like when all this started to happen I'm like oh my goodness maybe I was just too early and maybe we should try and give this give this another go.

>> And so that like that really grounded at least the motivation of saying okay I've done this before >> and like I've learned a lot over like my entire career of what it means to build developer tools and what it means like build something that can kind of span a lot of ecosystems. And so we started Gemini CLI from that point.

So, I was like I basically I went dark.

I'll say this. I went dark for like a week and um I'm like, "Okay, I'm gonna do this. I have this idea.

" Went dark for a week. Talked to no one.

Spent every waking second. Was getting like five to six hours of sleep at night.

And um this was like I think it was like Saturday through Sat Saturday through Sunday or something like that.

And then like that Monday I had a um a prototype.

>> Oh wow.

>> Did a video I put it out and Googlers started just going like wild for it.

It was super cool.

>> Yeah. Thanks so much for sharing that.

And so I guess building on that, you know, making the Gemini CLI open source was a very deliberate choice, I think, which you've linked to security and learning and a whole host of other things.

So how does this open trustbased approach contrast with the more like blackbox nature of other maybe major AI tools out there? What kind of ecosystem do you hope that this will foster, the transparency will foster? Yeah, like to to be frank like like I started my career in open source like I have always been in developer tooling or frameworks or one thing or the other and open source has been like front and center.

So like from day zero that was a key thing in my head but most folks feel like oh it's open source why would you not go ahead and do it?

Open source is not free. It's not free.

It's actually a very challenging thing to get right but oh my gosh it's so rewarding when you do. So like knowing this it was always this question like okay um let's weigh the actual balance like should it be open source should it not be and like the answer just ended up being like obviously it should be open source like especially with a CLI tool like this that's for like for a wide variety of people and of course developers as well but something that people could see and could understand okay it's running on my box it has a lot of capability so what does it do and can I trust it it's like that kind of like you mentioned And um the security aspect that was one of our biggest reasons for um like open sourcing this is we want people to see exactly how it operates.

Um so they can have they can have trust.

They know we're not doing anything behind the scenes. They know exactly what's happening.

And also means that if we make a mistake that we can fix it, that we have the entire community to help keep us grounded in what makes the most secure sense. We to be frank we struggle to keep up. like we we totally struggle to keep up with all the energy they give.

Um but to be frank, it's one of the most important things that we have.

And so when people ask me like what is the number one thing that's on your mind for Gemini CL? Like it's our open source community. Like by far it's number one.

Um and that's like I think we we've shut down the team like literally everyone on the team for like days in order to make sure that we can try and keep up with the amount of traction and the amount of energy coming at us.

But I think that's because we see the value.

We see how valuable it is to kind of build this in the open together so that folks have that that confidence that we're doing the right thing.

And we have that confidence that we're building the right things as well.

And it's funny, we actually still do the updates today.

Like we actually on all of our socials will post um every single week on Wednesdays, we'll post like 100 to 150 features weekly. U features, bugs, enhancements, all the above.

You said 150 features a week, right?

That's a lot. Um, what are the mechanisms that you use, you and the team use to make that possible?

>> Because we use it to build itself like it allows us to it allows us to do a lot more.

Like one of the coolest things I think that we've started doing is when you teach it to boot itself, it means it it can spin up parallel like threads simultaneously and each of those can go ahead and tackle different problems. Combine that with git work trees and you're going even further.

So far gone are the days of being able to ship like twice a year, which is a huge part of software historically because in AI every single week is like an insane amount of time.

Every single month is like Yeah, >> absolutely.

>> So yeah, I think it's it's honestly it's just that it's a mindset such cultural thing that we've built. And I also want to kind of acknowledge our open source community again as well, which is they really help make it possible because without folks really contributing, helping each other out even um I don't know how easy would it it would be like we put human eyes on every single one of the changes that go and we don't we're not just like letting things go without looking at them. It's just we have a lot of really passionate, really smart, capable um and motivated people to kind of make the right decisions at scale.

>> You use Gemini CLI to build Gemini CLI.

So, I'm curious to know a little bit more about how that journey started and and kind of what was your favorite part of it.

>> Yeah, it's kind of nuts. Like, I still remember actually the first feature it built for itself. Um, which is it's kind of crazy.

It was I say it's so long ago, but it really wasn't that long ago.

Um, but early on we knew that markdown rendering was going to be important, right?

So, I'm sitting there and I'm trying to build out its ability to render Markdown instead of just seeing the raw markdown in the terminal.

And there was a lot of utilities and frameworks to make this possible.

And I kept I forget exactly the wall that I was hitting, but I kept hitting a wall and using one of these frameworks that was like just a limiter. And that was like a hard stop for me and like okay like I need to progress I need to do something and I ended up asking hey like what are my options that was the first I'm just asking because I was at that point I was just asking it more questions to to learn more and it's like oh I can write a markdown parser for you said cool go for it and so this is the first time when it oneshot its own markdown rendering and to to be frank, a variant of that is still used today. Uh that actually renders.

So if you're ever curious, like if like Gemini CL CLI has written a significant amount of its own code, a super significant amount.

>> That's so cool.

>> Yeah, it's uh it we like I think the biggest thing that we think about on the team is with AI using AI these days, it is so easy to 10x yourself, which sounds crazy to say, but 10xing is easy these days.

100xing.

That's the hard part. Like that is where you really start getting into how can I parallelize my workflows to make sure my time that I'm investing into each of these things is best spent because models themselves, we talked about it there like there's so much that's left unsaid when asking a question.

They it still needs um human feedback a lot of the times in order to be effective.

Like things don't just get oneshotted left and right.

Of course, you see that in all all the videos out there and like all the highlights, but oneshotting rarely happens.

So, how do you optimize your time to make it so that multi-shot scenarios are they land super well and they can really amplify what you do?

>> Another thing I'm curious about is uh are there any like methodologies you think about when using Gemini CLI?

Like what's your mentality on how the tool operates?

>> It's grounded in a lot of my developer background, especially with AI.

um which is I personally feel like there's so much that's left unsaid when people will do a prompt. So like one of the big things these days is context engineering and it's kind of grounded in the same mindset which is when you ask a question to your to your AI of choice.

You're giving it as much information as you possibly can to enable it to derive the right answer. Now, of course, you're also leaning on its ability to like figure out more information and dig through your codebase um or whatever you might be doing at the time, but that alone is a really challenging spot because it can't know those offline conversations you have with your colleague or your friend. It can't know potent I guess it not usually does usually doesn't know your emails and your chat messages, right? I guess it could um but it doesn't typically know that.

And all those pieces of context really feed into building a coherent response.

Um, so like if you're currently writing tests and you ask it to go ahead and implement something, chances are you want to implement that in the form of tests, right?

Because that's your current your current workflow.

Um, that methodology though, it kind of rings true with every decision we make.

Uh, I think the thing I tell the team is that do what a person would do and don't take shortcuts.

So one thing that might be surprising to folks is we don't use embeddings.

Um, like for instance for search, we do not index your codebase.

uh we do agentic search >> which means how you as a person or you as a developer would dig through a piece of code we do the same thing we'll do find we'll do we'll gp our way through the codebase we'll open files or read them uh we'll effectively run commands like find all references on your behalf if we need to uh to find out what's the next piece of the puzzle because in the end we're trying to provide the right context to the LM so that it's grounded in every single thing it does uh to come to a good a good result. So I think I think it's probably it's probably the biggest one.

Um I think that kind of really resonates with us.

>> Something that I think is really cool is like maybe I'll ask the CLI to do something and uh it'll try to do the thing, but be like, well, I can't actually do this. Um but then it'll give me the steps that it needs to take in order to do it and be like, are you okay with me going ahead and doing these steps to do the thing that I need to do to get to your original request?

which is very cool.

>> I love that. That's that's like such a huge unlock is we call it self-healing, right?

Like its ability to self-heal goes so far and like when it it'll try things like it has a good idea of what's on your box and it tries to like use all those things that when it can't do it, it's so good at coming up with other alternates.

Like I recall this one scenario um where I was talking to our marketing folks and uh I was giving them a demo to show what it could do and and then their first question was oh can I like can you give me a link to this and I'm like oh um it's but we don't we don't have built-in like deploy functionality yet.

Like there's a cloud run there's a cloud run extension actually being released um very soon that will enable this just so folks know.

Uh but we don't have that built in.

And so they're asking how can we do this?

I'm like, "Okay, well, let me just ask.

" And what it did is it it it ended up like creating a GitHub repository.

And GitHub repositories have a thing called GitHub pages, which allows you to host static content.

And then it pushed this content to it and it gave me a link.

And I gave it I'm like, I had never even considered that.

And my question was, how do I give this person a link?

And it put all the other pieces together to kind of make that a reality.

>> Like the scrap scrappiest developer.

Yeah. I'm almost thinking of the use case where I'm like, "Hey, Gemini CLA, how can I like tell my mom that I'm still alive, like update her and it'll like ask me, maybe you should do that, but here are the steps that I would take.

" Like, "Are you okay if I open up your text messages or whatever?

" >> Multiple.

>> Totally. Yeah. And my mom watches these, so mom, I'm still alive.

>> Well, and also a moment ago, you mentioned something about like an upcoming feature on the road map.

Um, so what else like what else is coming up on the road map? Like are there any features coming up that you're most excited for?

>> Because we view Gemini CLI as being a lot more than just developers and because we've seen internally it's unlocking every professional whether you're a marketer, financer to of course a software developer. Um, it kind of attacks like hits all those those buttons to make it so we can work with every one of these professions.

We're really doubling down on extensibility.

So you can heavily extend Gemini CLI and this is not just MCP servers.

This is like literally you can install an extension which is a bundle of it could be MCP servers um specific instructions specific commands lots of different things to drive a different so this is the cloud run one that I had mentioned earlier they have an extension and this extension can be gemini extensions install and then you can pass in cloud run and it's a seamless installation um but enables you to really curate the experience to your like your preferences so for instance If you are like a Go developer and you want to make sure your environment is super Goofriendly, like you'll install the right MCP servers to make that happen, right? Or if you are it's a content generator, maybe you'll hook it up to um all your various socials.

Uh maybe you'll create like generative media APIs and you'll also hook that up to it. So being able to turn these on and off is something that's super important to us because we know that there's a lot of use cases.

And so the big so that's the biggest feature that we're going to be talking about soon which is how to build these extensions, how to install them, manage them um with the intention of u making this super seamless for people where people can spin up their own registries for they want to. We're going to have we're eventually going to have like a centralized registry for all of our extensions.

Um but the extension ecosystem is going to be the big one.

I can't wait to see what people build.

like we have a and we have a number of them coming out um from Google from from cloud um and just in general to really hook Gemini CLI into everything in a really seamless way.

>> Yeah, Taylor, I just wanted to thank you so much for sharing your insights with us and our audience of agent builders.

It's been a fascinating look behind the curtain of Gemini CLI and you've shared so much with us. So, just wanted to thank you.

>> Oh, thanks for having me. This has been an amazing conversation. Like I love I love diving in and especially love just sharing stories and uh if you haven't checked us out, check us out on GitHub.

Um you'll find us all the amazing like Gemini Google CLI on GitHub and then of course uh you can look for us on socials as well to get those weekly updates that we push out regularly.

>> Oh yeah, good plug. Yeah, thank you so much Taylor.

This was so fun.

>> Yeah, thanks so much Taylor.

And that's our show for today. Thank you for joining us for this deep dive into the Gemini CLI.

We highly recommend you try it out for yourself.

>> And if you enjoyed this episode of the Agent Factory, check us out next time where we'll continue diving into the world of AI agents. Until then, I'm EMTT Mirage.

>> And I'm Molly Pettit.

>> Powering down.

[Music]