Google has COOKED once more! Firebase Studio, Agent Space, & More!

As of late in the AI world, Google has been like this little mysterious chef up in the kitchen cooking up all kinds of delicious AI news meals. And today they're presenting us quite the smorgers board to look over and pick at. There's some pretty cool stuff, so let's check it out. First thing you need to know is all of this was announced at Google's opening keynote for Google Cloud Next. How does Google Cloud relate to AI? Well, I guess all of their AI models are hosted on the Google Cloud. So there you go. The first thing they're announcing here is Gemini 2.5, which we've already known about, we've already been testing and actually have had access to Google's 2.5 Pro for a decent amount of time now. And it's a phenomenal model. So this is basically just them bragging, saying hey, we've got a fantastic model. And that is absolutely true. Gemini 2.5 is a thinking model, as they say here. It reasons through its thoughts before responding. It's their most intelligent AI model ever and it is the best in the world according to the Chat Bot Arena Leaderboard. I do agree that Gemini 2.5 Pro is a phenomenal model. I mean you really have to try it out and in fact you can try it out for, for completely free. But this leaderboard right here, the Chatbot arena leaderboard, it's not my favorite way to benchmark how great or how good a model is. You can see the troublesome Llama four is actually up here in the second hand place. But that doesn't necessarily mean that llama 4 is actually better for a lot of use cases than even ChatGPT4O or Grok3. I get it though. Google's just trying to flex their muscles a little bit and 2.5 pro experimental is a phenomenal model. However, they are also flexing the fact that they have the highest score ever on humanity's last exam, which is one of the most, if not the most difficult industry benchmark. GPT 4.5 OpenAI's last non thinking model only gets a 6.4%. Deepseek R1 gets higher at 8.6. So thinking definitely helps on this benchmark. O3 mini bumps up to 14% on humanity's last benchmark, but Gemini 2.5 Pro experimental gets 18.8% humanity's last exam. I mean that is a tough benchmark and it does pretty good, at least compared to the other models we have access to now. There's still a lot of room to grow so this benchmark in my mind is a lot more impressive than the one that they show off up here. Moving on, Google has just released Firebase Studio. This is a competitor to Cursor Replit, Bolt all these different AI xcoding platforms. J Sensei says it's kind of like all of the platforms mixed into one. And Firebase Studio is available right now. You can get access to it. It's going to be linked down below. And I really like the fact that they have two different options right off the bat if you want to use this thing. First of all, we've got just a regular prompt box. This is for people that don't really know much about coding. If you want to just send it a prompt and see if it can build an app for you, this is the way to do it. But if you're more experienced, if you're a developer, you can start a new workspace. You can actually import a Git repo, you can actually get started with all these different platforms here, including Java, Python, Next js. It's pretty awesome. I'll do something really basic. Make me an app that develops a fun, secret spy movie style plan from any photo you send. There was an error creating the workspace. Okay, it looks like Firebase Studio is getting slammed right now because we're just getting consistent errors. Okay. I was able to manually set up a little bit of a workspace here though. It really looks like they're trying to automate the development process here with Firebase Studio. But yeah, look familiar, guys? Seems to be initializing things all by itself. It's already got code in here and boom, we're loaded up with Next js. But yeah, if we wanted to modify this, make the whole website green and lemon themed, this thing is preparing file changes now. Oh, wow, there it goes. Just spit something out, update the file, and it's updated. You can see it actually picked out a couple of different colors, yellowy and green colors. And there it goes. It changed the website to make it all green. Very, very cool. Obviously there's a lot more that you could do with this. I want to dive deeper in, but there's a lot of other stuff I want to talk about in today's video. This blog post, which will be linked down below, goes into a little bit more detail surrounding Firebase Studio and the AI interactions. So obviously what we just tested there was the interactive chat translates your natural language instructions into direct actions within the workspace. Instead of just simply offering suggestions, it actually performs tasks that you describe so you can Ask it to run specific test or extract a function into a new component. Try it out with all kinds of things. I'd be really interested on how far developers actually get to push this new system and how well it really works. We already saw this when we first loaded on our project that automates tedious steps allowing you to do lots of basic setup stuff. If you want to add specific things to your project, all you have to do is just ask the AI. It can generate files, install packages, configure the project as a whole. So you can just get straight into building without any of these slow, sluggish setup. They also say they're already exploring some new possibilities like multimodal interactions. Maybe we can actually send it a photo of a design that we came up with and drew out and it could build that for us. Yeah folks, give this one a shot. Let me know what you think down in the comments. I'm really interested. Okay, but that is not all that Google is cooking up. This looks like a nice little filet here. Google Agent Space. This is a new agent to agent Open protocol to enable AI agents from different vendors and frameworks to communicate securely and of course collaborate, actually perform actions and get work done all autonomously. Check out this demo of Google's Agent Space in action. It's pretty crazy, very chat GPT esque. Basically we're going to upload our file and our prompt. I am hiring for a staff software engineer position. Attached is a job description that's obviously in the PDF. Can you help me fill the role? Will the agent be able to complete this task? We click go. Obviously it's going to start thinking here. It's going to go through and look at the prompt and the PDF. So check this out guys. As you can see, it actually initiates an agent protocol. It wants to discover agents capable of candidate sourcing for the job position. And it already has found based on the available agents in the tool registry, the sourcing agent seems like the best fit. So it discovers the correct agent for the job and it's going to go over and actually call that agent and give it a task. It does actually ask a follow up question, which I love when these AI models and especially agents actually do stuff like this because it means it's looking for additional information to actually get the task done properly. I couldn't find information about preferred location of candidates. Is there a preference for candidates with any specific time zone range? User is going to go ahead and answer that question giving a specific time zone and range. But yeah, you can See showing off the collaboration of these agents, its ability to ask follow up questions so we answer. It does more thinking. This is quite interesting actually. It has a user experience negotiation within its own thought process. It can send the top candidates as just a plain text message or it can render them in the UX as friendly cards via iFrame. It goes ahead and checks to see if the protocol supports it and it does. So it decides to go ahead and use that nicer experience and boom, the agent has come back with five different candidates. That was in about 40 seconds. Seconds or so it got the candidates for you. I mean that is crazy. Now here is the crazy part. It's going to go ahead and create a reach out plan, but we're going to pause here and wait two weeks later once interviews were complete. As we can see in the same Google agent space interface in the recommended section, update hiring for a SAF software engineer can actually ask the AI for an update on that task. That was two weeks ago. So he's going to go over, click get update and let's see. These are the three candidates who have passed the interview and are pending a background check. So then we ask it to initiate the background check as well. I like that it even comes back with its issues too. Because it was successfully able to do a background check for one person, but it failed to do it for two others. We asked it to simply run a background check for the non US candidates. It contacts an international background agent to actually help complete this task. It says it's going to send updates with the latest status. Estimated time to completion is about two business days. To do these full background checks, which is pretty reasonable. I mean, what a capable agent. At the end of the day, this is the real big kicker here. Task and state management. Clients and agents need mechanisms to communicate task status changes and dependencies throughout task execution. This thing can take a pause and say, hey, I've sent it off to this agent. We'll hear back in a couple of days and all you have to do is log back onto your Google agent space and get an update. This is pretty wild. I've never seen an agent quite like this able to hold on to tasks and complete them over time, like long periods of time. It's a pretty big deal. And they also have a blog post on this one. Although this is a little bit more about the protocol, the agent to agent protocol itself, which is allowing the general capabilities here that we just witnessed. So this protocol is open, which is great to see. Obviously they want people to use it to get started, it's a quick PIP install but it's their agent development kit. So for you devs that is going to be linked down below right now it supports contributions from over 50 technology partners. Some notable ones in here, I see Cohere, Intuit, LangChain, PayPal. So definitely some pretty big names in there. So this protocol allows the agents to communicate with each other, securely exchange information and coordinate actions on top of various enterprise platforms. I like the design principles that they lay out here. These are key in my mind to actually make these agent to agent systems work effectively and allow people to kind of just get up and go and use them. First of all, embrace agent capabilities. So they need to be able to collaborate in their natural unstructured modalities even if they don't share memory tools or context build on existing standards of course, like HTTP, they claim this is pretty secure enterprise grade authentication authorization. Hopefully this is the case support for long running tasks, you know, things that are actually useful. So this is flex to support scenarios where everything can be done quickly or something that might take a little bit more deep research hours, days to complete. Finally, they want to support a bunch of different modalities. They don't want to just limit it to text, they want it to support audio and video streaming as well. So pretty great list of design principles for this. It looks promising. It really does. How's this sucker work under the hood? The protocol facilitates communication between a client agent and a remote agent. The client agent is responsible for formulating and communicating tasks, while the remote agent is responsible for acting on those tasks in an attempt to provide the correct information or take the correct action. This interaction involves several key capabilities. First of all, they can advertise their capabilities with an agent card. So that allows the client to identify the best agent for a given task, which we saw in the demo. Task management. Obviously the client and the remote agent have to have good communication to end up fulfilling the request. At the end of the day that kind of goes along with collaboration and then obviously obviously the user experience negotiation, how will we present it to the person who asked for the task to be done? Overall though, this is pretty darn cool. A nice open framework for us to build agent workflows on and it seems to genuinely be quite capable. It's good stuff, Google. This is a nice, nice treat. And finally, our last little piece of Google News is that they have officially released VEO publicly. It's generally available. It's a text and image to video model. A lot of people say it's one of the best, if not the best in the field right now. And a public Preview of Gemini 2.0 flash is also available. I've actually already had access to VO2. I want to dive a little bit deeper into it, so maybe I'll spend the rest of the day messing around with VO2. But Google had some sound updates today huh? Starts off with them bragging about their latest large language models, but moves into some really interesting tools for developers people to create apps, use agents to complete tasks. These are big use cases right now for AI using IT to code, create, develop apps using it to complete tasks. More complex, actually difficult ones simply by just using agent to agent protocols. I'm excited to see where all of this goes. I really am interested to see what you guys think about Firebase Studio, so let me know down in the comments. But yeah, that's going to be it from me for today, so thank you so much for watching. All links are down in the description and if you want to stay up to date, really be on the cutting edge of this, I suggest checking out my Discord server. There is a great community on there and you can share all the different things that you make or do with AI. Also check out my Twitter. Anyways, yeah, I'll see you in the next one. Thanks for watching.