Meet Jules, Your Autonomous Coding Partner! | Oleg Aravin, AI By the Bay25
Recording: Meet Jules, Your Autonomous Coding Partner! | Oleg Aravin, AI By the Bay25
All right, welcome everyone. Thank you for joining me here. So, I'm Ole. I'm presenting Google uh autonomous agent named Jules. Uh yeah, I'm staff software engineer working for Google Labs. Like if you didn't hear about Google Labs, it's a little different than just like overall Google. We're trying to make things like faster with less like bureaucracy and just like with startup feeling. All right
Before before I start talking about the agent, let me ask the question like who is actually like how many people here actually writing the code? Like raise raise your hand. Okay. Some some people how many of you use coding agent to help you with your work? Okay. Okay. I I I think it's more people who are than actually writing code and who are using jewels. Yes. [laughter] [gasps] Thank you. All right
Uh before we start talking about AI and agents, let me take you a little bit back in time. So, and tell you tell you some personal story. I've been writing software like coding for like almost 30 years. Like I wrote my first line of code when it was when I was like eight or nine years old. If you remember that time it was pretty like interesting just a black screen like white text like nothing and your compiler your best friend who is always telling like how wrong you are when you're doing mistakes. Um and then fast forward to like 2000s we got some modern ID. We got syntax highlighting some like code generation like generating Java setters getters automatically that felt like amazing right some like new generation you can you can do things easier then going back not going back going forward like 2020s like we got the rise of AI right first time we started we had ability to generate some code snippets with just a prompt right I remember first time generating some function And with just a prompt that felt like absolutely amazing, right? It saves you like so much time. Now 2025, we are in a like time when autonomous agents are coming and rising
Autonomous agents, they do all the work for you. Not not just the coding but the complete like feature set including code test like and everything. [clears throat and cough] And um with that I want to tell a little bit how different like uh autonomous agents and uh like just ID copilots which we used to use sometime before. Uh so with copilots they pretty much live in your ID. They help you to write the code. You can see the code. This is pretty interactive. You write line by line
It's kind of like you you still the driving uh but like the coding partner kind of like assists you with the coding. With autonomous agents the f things are different. They live somewhere in the background. They work for you, right? You don't usually see the code. You may not even see that, right? It is full as synchronous. It's not interactive. It's not like line by line. This is full interactive
And the result of work is usually big. It's just it's not just a code that you can modify but it's some complete thing like a pull request you can ready to merge and you super you pretty much you supervise so you don't drive but you supervise this is the main difference [clears throat] and with that I want you to meet like Jules [clears throat] this is Google a synchronous like agent it's uh developed as a synchronous with a with a minds uh like it was initially designed as a synchronous agent. It wasn't evolved as some like ID helper uh into something bigger. So we designed it uh with this mindset how it can be fully autonomous. How it can help you to do things like features for you without you actively engaging with that. Uh it comes with a lot of like features like coming from this mindset like planning. It does a lot does planning. It doesn't just come up with like coding for you when you ask, right? It it it has a plan you could which you can modify and you can adjust
It's full autonomous. You just give it a task and like fortunately like in ideal case you just forget about it until it is done and you come back and you see the result of the task. Uh it has a bunch of integrations. We have integration with GitHub like more integrations coming like Slack. So like ID you can use it from wherever wherever you want wherever you you're working in [clears throat] and um this is developed by Google which comes with a lot of benefits like privacy and trust it runs in a Google environment so with so it is as like as good as Google infrastructure could be like as secure and of course it's powered by Gemini 2.5 which comes with a like large context Um so like le let's let's be honest like here for a bit it's 2025 like and everybody is doing agent right if you're a developer and you are not doing agent today you either doing something very unique or or you're doing something wrong right >> so and here I'm presenting you yet another agent that you can use and you ask a question like hey why do we need this yet another agent. How do I compare it to what I already have in the market? This is why I put this slide here. This is like it's not a full list, but this is just to give you idea where Jules fits in the current market of agents. [clears throat] If you see on the far right uh uh like in the landscape of fully asynchronous autonomous agents that's this is where Jules is in with Devon Copilot agent with codex if you use it in like agentic way uh [clears throat] and you know like the borders are like shifting right even not a synchronous agents right now who are acting like assistants for you they're trying to get to the space of full autonomous Right? Because the context get bigger and you're kind like you as developer starting to have less and less control over what you're doing
Right? And this is probably what you you will be doing in the future. Could be, could be not. [clears throat] Uh how Jules is different? Uh you may ask this questions. Why would I use Jules? There is already Codex like Gemini CLI on the market. What is Jules? What what is so good about Jules? Uh there are there are a lot of differences but I wanted to highlight just a couple of them. As a synchronous agent, it's always on. It's run on cloud. [clears throat] It's there for you
It doesn't sleep. If you suddenly have a task on Friday evening, you can just give it and come back Monday morning and it's it's going to be ready like fortunately. Um so this is obvious which is coming with with it being fully asynchronous and autonomous. Another thing it's built for scale like as I mentioned before we designed Jules with a scalability in mind with a uh with a mindset how how big of a context it can take. It can take your entire repository and put it just in the context like thanks to Gemini for that. Uh [clears throat] so you can it shines where you when you give it like really complex tasks not something oh just write a unit test for me or just make this function but to complete the entire feature like if you have a specification of a feature you can just give it to Jules and Jules will implement everything including the testing all the the feature front end back end like pretty much everything the more complex task the better Jules performs of course it's going to perform on a small task as well. But the bigger the task, the better looking you can the better you can leverage jewels. what Jules is and under the hood we have a lot of like novel agentic like workflows uh like some of them like we have like we like good planning right that you can modify we have different sub agents like one of them is critic agent when Jules does something does doesn't doesn't just like give it to you there is another agent coming and which is called a critic agent which is trying to figure out was it the actual original task right? Was it implemented right? All the testing is done
So and all all just all after that time when everything is passes and all sub agents kind of agree that this is good enough you get the result. [clears throat and cough] [snorts] All right you ready for demo? Do you want a live demo or a recorded video? >> Live demo. Okay, risky live demo from the stage. Something may not work. So just keep that in mind. Okay, so this is Jules. [clears throat] This is uh this is front end. It's super easy
Right here I'm already connected to GitHub. If you just come here for the first time, you're just going to see the button connect to GitHub. It will install it. Jules will be installed in your repository with a bunch of permissions and you'll be ready to go. You'll be able to choose a branch and repository and you just give it a task. That's it. Like no virtual machine environment configuration is needed. So this is intention that you can start like super quick
If you need to configure it later, yes, of course you can do it. So let's test it out with some simple task. So I built this website for this like demo purposes just specifically. So I basically I put the entire presentation into jewels and ask it to come up with some website right something I can modify. So I'm not doing it I'm not doing anything you see I'm just using agents and uh and Gemini. [clears throat] So this is pretty much everything you see in the presentation. And let's give it a simple task. For example here I want to implement the dark mode like toggle right something simple
So I go here and I just type it. Right. That's it. Uh so then I can choose the interaction mode. Uh default is start. It's basically you start and it does everything for you. We also have interactive plan. If you want to engage with the agent, you want to learn, you can ask it more questions, you can still do it
Uh here I just click start and that's it. So the task is getting started. So what it is doing right now, it is booting the virtual machine somewhere on Google environment. [clears throat] Then it's going to clone your repository and execute the code change. It's going to take a while. So we will not babysit here, which we don't supposed to babysit, right? We supposed to come here when it's done. So while this is working, let me just walk you through some other capabilities of Jules. Uh you have a list of tasks like here on the left
So you can execute them in parallel. uh by default you have a so like I have pro subscription here but by default you can execute 15 tasks which is plenty right because within one task you can do large like thing and so we're not limiting right now by the number of tokens and by the complexity of the task but if you want more you can up you can just get more with Google one subscription uh we have some settings some conf something you can control all uh your export settings like pull request more integrations coming so we just built this integration for Slack it's not public yet so you you see it for the first time uh so ability here would be you can connect your slack workspace and you can kick task right from from uh from jewels so let me give you example of the task how I created this website site that you've seen I just give it a des I just give a description which is like a full spec [clears throat] what what should be done here. The more detailed the better. Uh I can even attach pictures. I can use I can say like please use style from kind of like this picture or the color schim and it would understand it. Uh it came up with a plan. You can see it's pretty structured. I know the text is a little blurry but uh you get the idea right you uh this is a plan that you can modify if you realize there is something else needs to be done you you can you can do it here and then it starts reasoning it is saying what it is doing like how it is doing it gives you all the comment it executes like if you see like it came up with a python file I didn't tell it to come up with a python right it should figure out come up with Python
It tell it tests it. It makes sure it compiles. It doesn't give you it never gives you something that doesn't compile, right? It can give you something that doesn't build on your environment, which is like future step of integration. But so far like our the code quality is our like main priority here. So we're trying to like honestly like you're not going to use agent if it doesn't come up with a code that you need, right? Uh this is essential because everything else is just built on top of that. So it knows about pictures UI you can see all the screen you can you can tell we have memory you can see here. So during each like task execution it's trying it is trying to figure out what is the important information you have about the repository or about the task you are giving. So and it's it it tries to remember it
So it puts here and you can modify of course you can modify or you can add memory and this will be reused during all your next like task executions. So what else and you see like at the end it just comes up with a pull request so with a description and it publishes a pull request and this is what I did and uh it doesn't stop here. uh if you say so when I have this pull request I realize okay I want to have some modification here I [clears throat] didn't like something and if I go back here to the session to the task I see there is some like logo placeholder placeholders like left to to Google sign I didn't I didn't want it I wanted to be removed so how would you do it like you can you can remove it from the code by yourself or you can just ask Jules to do that you can ask it here like directly in the UI by something typing here or what I what else I did I just went to the pull request and I put a comment here in natural language hey I don't like this play holders please remove it and it did so and it came up with a new comment and uploaded automatically to the pull request so let's see how the our task is performing uh you see it is working right now. It came up with a new UI which is like in dark mode and this is still like working right. So we'll we'll come back to that later. Um uh you can start session differently. So here I just want to show you different capabilities. This is GitHub and uh I have issue here
So for my website like let's say like simple thing I want to implement some winter theme like we know winter is coming right. So let's uh let's build this functionality. How can I kick this task from Jules? Easy. I just put Jules label here and let's see if it works. It should come up with a task. [snorts] It did. Okay. You can see the result here
If you go back to UI um like refresh it, you see there is another task just started with all the description taken from the taken from the GitHub. So when it's done, it will come up with a pull request will update task. So the idea here you can engage with Jules on the web if you want or you can engage with Jules from the surface you're working on. If it's Slack, it's going to be from Slack. If it's GitHub, it's going to be from GitHub. Uh so you see it came up with a plan. So let's go back and see how our task is performing. It's taking a little longer than it you it should be
So it's supposed to come um come back already. [snorts] Um uh what else can I tell you about Jules? So like another capability I didn't mention here. Uh we have all the coding here like if you still want to see the code you can see the code right. Uh if you want to you can even download like zip file of the code if you want to continue working with your like environment uh like somewhere uh we have CLI as well. We have CLI we have API you can build on top of jewels if you want to uh so you can customize the flow in the way like in the way you want. Okay. Um what else? Okay, we'll take questions uh after that. Uh so let's come back to this uh [clears throat] task a little later and let's go back to the presentation
Let me just have one more slide and then I'll come back there. So like you seen like jewels in action like right now all the current capabilities like large code changes quality code quality like memory management like deep GitHub integration um in the future it doesn't stop here of course in the future we want to continue exploring this mindset how it could be more and more uh helpful for you right now it's more right now you can think about it as like junior developer You give it a task and it comes up with something. But it's small. It's right now it's like software developer who you give it a task, it comes up something and then it just sits and doing nothing, right? It's not happening in the real world. So we realize it and we want to build some proactivity uh into the jewels. So based on the information it knows it will try to come up with some ideas suggestions to you. If it's see like to-do that is not resolved, it will come up with a task for you to resolve this to-do. If it see vulnerability, it will try to resolve it and give it give it back to you
We want to increase uh support of large repositories. Right now it's pretty much tied to one repository, one branch. So we want to make it work for across multiple repositories, across multiple branches. Like you you may commonly see the case where you have front end and back end in different repositories, right? If you want to come up with a large feature, usually it goes across multiple repos. So we're going to add this support for that as well. And of course we are like doing we'll be doing more and more integrations from some s bug tracking systems like linear like Jira of course we're going to build MCP. If you're doing agent and if you're not building MCP like again you might be doing something wrong in the modern world. So MCP will come u okay let me stop here for a second and uh let me let's see if it came up with a result
Okay it didn't come up with a result yet but you know this is live demo right? This is what you asked. Um, but as I see this task is actually complete the winter theme and something we kicked like a synchronously from the GitHub and the pull request is here. So we can see it has a description and uh here I have this uh site developed in versal and if I click preview I here we see yeah it's a winter film it is generated right [laughter] [gasps] and I did nothing right I just made a prompt I just give it a task this is like amazing like if we think back about like the timeline I was talking about before right you do the coding changes and then it come up you can do the function as a prompt now you can do the whole thing right as just a prompt this is amazing like like I'm develop I feel like super excited about that like [clears throat] okay back to presentation uh so you've seen jewels you seen you seen it in action uh it's available right now you can use it uh thank you for using it already. So you can tell how great it is to to the audience. Uh so it is available right now. It has a pretty generous free tier. Uh you can use it from the web, you can use it from API like CLA wherever wherever you want and you'll be able to use it from more surfaces and um like question feedback. We are like on discord, we are on X, we like as developers we like we we engage with people directly which is not common for Google usually only marketing people like interact with real customers but we are interacting with real customers and like I'm here as engineer like interacting with all of you which is great right so you can give feedback directly to me so at like at the end before questions I want to get little like philosophical, right? Um those agents like autonomous agents, they developing so like rapidly, right? They do a lot of work for you [clears throat] and like the question may come naturally is it going to replace me like as an engineer like what what we're going to do is like engineers in the future sometimes even for like myself I got like thinking what I'm actually doing like I'm engineer I'm writing a lot of code I'm like five to 10 10 times more productive as I was like five years ago and it's not because I became five times smarter Right
I wish. Uh so it does a lot of work. So how do I feel? How is my future work is going to look like um so we as a team like and I personally see this uh it's going to more look like collaboration. You collaborate with an agent, right? It doesn't replace you. >> It just uh you supervise the work. You give it more work, but you have more time. you still have time for uh creativity for design for like organizing the system like figuring figuring out on a larger scale what needs to be done. Uh that's why I feel excited about it and I hope you too
Okay, this is all for me. So we have about four minutes for questions. I guess my question is like what is the most >> we'll bring the mic to you? >> Just raise your hand. [laughter] >> I was curious what's the most uh sophisticated thing you have built with jewels. >> Oh okay. So I can tell you about the time I've seen tasks uh that are running for like five to six hours as complex as they are. And so I personally I personally doing you know more like a smaller tasks with it like for demo purposes. Um um but for example like my I have my personal project and uh which does some like tracking system right tracking system for sports for myself which connects to straa gets all my activities so I can see some results there
It's a complex system which also works with like overflow like with database management with like it's it's a lot of like moving pieces and I wrote this I wrote like zero lines of code by myself it's all done by jewels iteratively not as a single shot right because like more I don't know what I'm doing completely right so I'm coming up with a structure I see the result and then I see oh I need this feature and I keep asking it to implement more and So but it shows that it can work with complex systems. >> So just follow up is that when you iterate is it giving you big PRs or is it giving you >> depends on complexity. It can give you really really large PR. It could be Excel which might not feel great right because how you going to review it but it can depending on complexity like here the CL which I presented here it is small right I didn't give it a complex task but this is for like time purposes right u but yeah to like directly answer your question it can come up with really large changes large complex and try it out it's like so easy to to do So I'll be like interested to know. >> Yeah. Thank you. >> Hi. Um thank you for the great talk
I was not aware of Jules. I wanted to ask it seems like Google has at least the Gemini CLI code assist and now I'm aware of Jules as well and as of this morning presented earlier today anti-gravity formed by the scraps of like uh well the I guess the people they hired from Windsurf >> right >> so like the question is are those competing officers with competing uh offers within Google like is Google planning on consolidating all of these into a single product >> or what is do you know what the play is here? I wish [laughter] I don't like simple answer I don't know uh but I can tell more context that all those offerings are different right Gemini CLI it came more like a CLI right which is coming to the agent more to the world of more autonomous agents Jules was built differently like full autonomous agent with complex tasks in mind we don't even give you coding right now we don't give we you can see the code but you don't modify it we're only thinking to build integration right now with ID where you can modify the code. So different mindset anti-gravity is also different. It's more like ID, right? Where you can use different agents, right? And they have their own and uh and it's all interconnected. You can kick tasks from Gemini CLI now. It was just built recently. So it's integrated with anti-gravity of course like we're going to integrate sooner or later as well. So I would say I'm hoping those all those offerings going to survive right because there is like no consolidated solutions that works across the surfaces and maybe it shouldn't be because it's just they're just different we're thinking about large scale like anti-gravity is thinking about different thing right how to present you with ID how to how can to how to help you modify the code in the best way you can modify the code but we want you not to see the code like [clears throat] at all
Right. So you don't you don't supposed to see like if the quality is good enough. >> You mentioned the generous limits on the number of tasks that run simultaneously. How practical is it to issue say four or five tasks have them all run for an hour and then come back with conflicts. >> Depends on your task, right? Depend depends on what you're doing. If you are actively starting the project and you're implementing a lot of features and maybe not interconnected maybe you want to do something in front end or in back end you just give tasks and that's it. So for example, how I work when I do things, we have jewels like internally available. I just run up to five tasks and just I just going to get some coffee, right? And I come back and I see the results, right? I would say it's practical, but it depend your focus mode where you need to crack some like one specific problem like you may not want to run it in parallel, right? uh you may want to like sit and babysit and interact with that
So depends on your task but for me personally I think it's really really practical. >> Thank you everyone and thank you Alec. Please give him a round of applause. [applause] >> Thank you. >> I know I know there are more