Developer Spotlight: The State of AI in Workspace Development

Key Takeaways

Maybe Justin, you can mention what your flow looks like when you're developing. >> Oh, it it's always changing it seems like right now, but it's a it's a lot of Gemini CLI to assist me and, you know, often, you know, an IDE and kind of back and forth with that to, you know, work through whatever problem it is. Sometimes I'm, you know, in an app doing deep research to generate like a markdown document and sticking that into my codebase to kind of guide things along. Welcome to the developer spotlight. Today I have two guests uh both in from my team. On my left is Justin. Uh right is Peric and I'm Vine. We're all part of the Google Workspace developer relations team. And uh yeah uh so let me just get started. Let's start with uh maybe Peric. Tell me about the event that happened the last two days here in Paris and then uh in Sunny Bell as well. >> Well, Sunny Bell was two weeks ago. Uh it's was pretty good. I I loved it. Um we had a really good audience. Uh good talks. And this week it was pretty similar in Paris. We had the same tracks, same talks as well, pretty much. Um and I had the chance to be the MC of a specific track uh which is the um it's called the in-house developer track. >> Okay. >> Uh the idea is to uh discuss uh things about uh developers who do things for their own organization or for themselves or for a specific team. So it you know will be mostly about appcript and customization things like that. Um, I think because I was the MC, I was able to see, you know, the this full track in Paris and I felt like it was mostly about AI, uh, OP script obviously because of the track, but a lot of AI obviously. Uh, but this time much better than last year I think because it was more about um tooling and productivity with AI, not only just generating stuff, >> right? >> Um, and I think Justin has a few things to say about it. >> What was your experience like? Yeah, it was just a lot of fun to hear how people are using AI to kind of be more productive with workspace and build add-ons and use appcript and everyone has a different pattern it seems like and so just hearing from everyone is great. Um >> yeah I think >> myself and everything >> right I think even so we all in the we all de developer relations so we do a lot of development like writing samples and documentation but I I feel like all of us even being in the same team we have a different workflow uh maybe Justin you can mention what your flow looks like when you're developing >> oh it it's always changing it seems like right now but it's a it's a lot of uh Gemini CLI to assist me and you know often you an IDE and kind of back and forth with that to, you know, work through whatever problem it is. Sometimes I'm, you know, in an app doing deep research to generate like a markdown document and sticking that into my codebase to kind of guide things along. Um, it it varies. >> It changes. Yeah. How about >> every project? Um, I'm not that advanced to be honest. [laughter] Um, Justin is really way ahead of me. Um but that so that's actually good because on my side um I'm still doing the you know basic thing of copy pasting from Gemini like going from you know um >> tabs in my you mean the Gemini app. >> Yeah. Yeah. So I will have like multiple tabs on my um on my Chrome and then I will just copy past you know like that and have you know building the context each time and things like that. But so I'm really really at the beginning. Okay. But the good thing is that um I learned a lot during the the summit because of you know just in talks and different things we um we've done um like the CLI and everything. I've seen that um live uh with your dele uh and I I I learned how people are trying to do it. So I feel like next step for me will be to try it a bit more. Uh I feel like people are more productive when they do that but it takes some time to uh adjust and uh find the right way of doing it. Uh I think it's different for everyone >> right. Yeah I know I I wasn't I fell into the bandwagon too of using Gemini CLI. I'll talk about that in a bit. uh but one of the interesting things was how even for myself I've started seeing how my developer journey has changed too u where I was never using AI tools and I'm using too much of them uh I know right now like uh >> how do you know it's too much [laughter] >> well well we have we'll figure that out uh but for example like one one of the talks that uh Justin you did about uh using all these tools U like MCP servers for example one of the problems I had was u Gemini CLI right now doesn't have a way to pass temperature and I was too lazy to you know submit a pull request and I was like I want to do this quickly >> uh so guess what I did I used Jules anyone used Jules >> um and I asked Jules to write the code for me and create a PR for the so it's still pending but yeah it did the work for me with >> it went from a feature request to a PR Exactly. Yeah. Yeah. Yeah. [laughter] Yeah. Yeah. Uh uh so that was amazing. Uh uh so yeah there's a lot of cool stuff around developer productivity. I know we had uh some folks that gave a talk on developer. Jasper's talk was one of them. Yeah. Unfortunately I wasn't in the room so I missed it. Maybe some Did anyone of you catch some interesting points? >> Yeah it was is interesting to see the layers that you know developers have built on top of these tools. um you know having sub agents and different personas and having this whole team of AI kind of agents working together to work on an app for instance um >> right yeah yeah and I know you've been working dabbling with some agentics ADKs and >> yeah I I feel like my feeling when I heard Jasper's talk was um he built this whole system only with LM >> okay >> and on on his own like you know with his own ideas and own concepts >> but it was something that will come sometime and being more toolled. And this is what we call agentic apps and you know this uh ADK AI agent for example uh nowadays >> which uh which is like you you have a root root agent but it it can play with a lot of sub agents that you can define and they will have like really specific purposes >> almost like minions like the the agents and minions they do. I guess the main difference is that minions are look pretty look pretty alike [laughter] >> but this time it's like you can you can specify will really specific um purposes and you can specify also the tools they have access to. So they are willy you know into a willing one specific thing they do pretty well and walking together they can they can interact with each other pretty well >> right >> and then come back with an answer that is like yeah uh that's a full chrome full full team uh walk delivering on something that is complex >> right nice >> so it's uh you're really going from I'm generating this using LLM to a lot of people or virtual people agents uh working together to produce something >> yeah But then there we have uh this problem where there's a lot of context that needs to be shared. I don't know if you have faced into a situation where one agent or one tool has no idea what the other tool or agent is speaking. How have anyone come up a way to solve this? >> Well, there's definitely protocols for doing that. Okay. Um I wonder if like this is a shortterm problem where like at least for like developer tools if you know we're just trying to work through some of the limitations by kind of creating these specialties if those will just kind of go away as the tooling and the underlying models get better. But yeah there in terms of like using that in your application there's standardized protocols that are being developed to kind of you know communicate >> right between these So you built one for the developer documentation for workspace stuff. Um maybe you can talk about a little bit about that. >> Yes. So that's uh you know using MCP or model context protocol. Got it. >> To expose you know tools, resources and you know prompts that you can share from uh kind of a through a standardized pattern um that works in you know different AI kind of applications and clients. >> Nice. And so for this MCP server that I created was for workspace developers to get the you know up-to-date documentation find code snippets to kind of improve the context for working with workspace. Um >> nice. Yeah. So I I had uh one so the Gemini CLI uh presentation I did for the Google devil developers group uh that was interesting. It was very wellreceived. I had a half an hour to do that, but we went over because it was so interesting, challenging, and fun. Uh, so this was my first time using Gemini CLI to build something uh that I can build on my own. I know where how to build a workspace add-on, but I had to show, you know, folks that, hey, you can write zero lines of code and build a workspace add-ons. uh and one of the challenges I had was uh depending on whatever model you do it hasn't learned new stuff. So you had to provide a lot of context in different ways. One was like MCP servers, the other was like the Gemini markdown file, right? Uh and so I'm also learning. Are you learning any new ways to make the AI work better for you? >> Not really to be honest. [laughter] Uh again I'm not really uh into all this research uh at least not now. Okay. But uh yeah >> uh I know you had a talk about uh developer productivity. Maybe you can share some stuff that attendees found really useful. >> Yeah a lot of it was just around managing context like you were saying about you know having the up-to-date information and kind of like customizing the context for the specific project that you're working on. >> Yes. um giving those hints and you know kind of asking nicely to do it a certain way for example >> right I know me and you had a discussion where we're talking about how um in short term we can try building an app just by providing the agents the Gemini CLI context from PRD is uh the product requirements document and design docs we haven't tried that have you tried that or >> yeah I did like I was saying a little earlier um taking you know like communicating with like Gemini app to like generate a PRD and then going from there to generate a design doc and taking that as markdown sticking into the codebase and then you know having Gemini CLI kind of use that as the background to like implement the app. So it's you know kind of working through some of the logistics that need to happen for the implementation right >> and giving that >> in [clears throat] that rules file or you know the context file >> right uh yeah >> so I guess I guess one question is because you you looked into it u if you had to list the most useful type of context we can provide you know for the the Google Workspace developers today uh and maybe we know will come in the future uh What would you list? >> Um >> I guess documentation with your MCP server for sure. >> But yeah, you know that will change too but yeah from uh >> well I'm thinking for example code code samples is something that is challenging to um to put in context right now. Is that >> that's right. So I'll go back to the Gemini CLI vibe coding session I did. Um, I was trying to have Gemini CLI create images a cat image using nano banana which is I think Gemini 2.5 flash image model and interestingly the Gemini model the Gemini CLI model hadn't learned on how to use this new model >> so when I was trying with it it kept using very like in the beginning was using deprecated models like flash uh I mean Gemini 1.5 five and I had to keep telling Gemini like hey this model is deprecated use new one >> and interestingly when I asked it to use uh the flash image model it said like oh the user is incorrect this model doesn't exist uh and had to provide it context like it does exist I had to provide it articles provide more context >> uh I even had to point it to code samples that I had in our GitHub repository >> so you knew the solution is like >> yeah like I was nudging it to do what I wanted to do. Uh yeah, but yeah, this those are some challenges. I know uh you had some challenges. >> Well, yeah, getting this back to workspace, maybe it's like uh you know, saying this is a editor add-on versus a you know, a different a workspace add-on that's used in the card framework. um you know putting those kind of that information up front in that context file can help things you know avoid going down the wrong path basically with the AI >> um >> any new features workspace related anyone excited about I know flaws is a big one >> I can I can maybe uh change the past a bit and stop talking about AI >> yeah yeah yeah let's do that >> so something we released a few months back I believe it was a few months back it's um you being able to uh publish a Google Workspace add-on with a chat app inside bundled u as a single add-on. So you would you would develop your add-on before let's let's take a step back before that you had to develop your chat app it would be um specific item you you will just have on the marketplace same thing with the Google workspace add-on for different uh applications like docs or you know slides um but many people were kind of upset or not really happy with that because then you have to u manage two different entities one for the chat app and one for the others. Uh so what we've done is uh provide a different uh a new framework um that makes it possible to bundle the chat app with the other apps within the same add-on, >> right? >> Um and that makes a huge difference because um then you can reach a lot of people using chat and it's also a new way to interact with your add-on which is in more natural form, >> natural language. Um, and then that's also pretty important today because we are talking about all these. Oh, sorry. I'm talking AI again. [laughter] >> Oh, no, you have to. >> I did two minutes without AI, but I'm coming back to AI. [laughter] Um, so yeah, the thing is now we have this um this chat app um that is kind of part of the add-on um the same add-on. Then you can also interact with natural language. And because you can do that, you have obviously the link to agents or AI agents that is quite natural. Uh because then you can talk to your agents about something >> that is related obviously to your add-on globally. Um and do something with your natural language instead of clicking clicking clicking uh on the add-on um sidebar for example uh when you're trying to do something. >> Yes. Yes. Uh so that's that's >> you built something around that right where you have um a chat app agent I don't know if it was a as well >> that knows about all your workspace tools. Correct. >> Yeah. So something I I didn't recently was playing around um with this because now that we have that uh I try to build an having an agent somewhere that is uh built for specific purpose and I I've been using ADK for that. um which is a framework to build agent right so it's really >> not Google workspace at this stage it's really something that you do on your side uh you are an AI expert or someone who who is able to do that you want to play with it >> uh or yeah so you have your agent uh but this agent is difficult to reach from Google workspace because we don't really have anything to include an agent in any way by default >> um so my thought process was well now that we have this chat app and we can you know have this uh integrated in a in a single add-on um that will be nice to connect to this agent and uh see how it goes. So what I've done is um I just wrote some code to interface easily uh with uh this ADK agent >> right >> um I deployed it on Vertex AI. So Vertex AI provides you with a few APIs that you can you can use to interact with this agent. >> Okay, nice. Um and then the idea was to make sure that um this interaction makes sense uh visually in the graphical graphic uh user interface. Um so what I've done is uh make sure that whatever the agent is um responding >> can translate visually on uh the chat app for example. So it will be like oh the agent is sending a message with the uh maybe a part of the answer because you have multiple agents working together. And sometimes it takes time. So you want to stream the responses as they >> have it on chat. That's nice. But what about having it in other layers, other apps? Um so that uh you know you can use this agent from docs as well. >> Uh so what I've done is well the API calls everything is the right but the UI is different uh because it's not a chat. Um, so I built something that is generic enough for all the different uh apps uh that we have in Google Workspace uh add-on and uh there it's you know it's not a chat but [snorts] uh you can see the the responses going on and uh being updated live right >> um and the good thing is that you can provide context to your agent because you are in docs >> so you can have access to your documents right or if you're on drive you have access to uh the different uh documents as well you have on drive. M >> so you can select something you you have your context of you know whatever you're doing or maybe you open an email >> and then you can just say okay I'm talking to my agent or I I make a request I want you to know about this email I'm currently selecting and then you know the integration is starting to be really interesting >> right uh but I know one of the challenges we had so the language of AI is text or markdown >> and but chat app the language is not essentially I know we had a new feature and now do you want to mention that too? >> Yeah, that's that was really an issue um to be honest. Uh this so this is uh I think we launched that uh one one >> maybe a few weeks ago. Yeah. >> So I I wanted to have something for the you know for the the Sunnyville Summit >> but I was like this is not you know it doesn't >> it produces Macdown but I cannot really have that in the in the chat. I was like, I have the whole thing, but it doesn't look good. Um, and it's it's I don't want to play with, you know, transforming markdown. >> Um, so I knew about this uh this feature that was coming up. So I was waiting really waiting for it. And uh and the good thing is that it happened right before the the summit in Paris. >> Uh so I was able to integrate that and that looks awesome. And there is another feature that uh that was um released at the same time. It's called um cowsell. >> Okay. Um so the good thing I believe it's only specific to chat but uh the the thing is when you have a list okay that's usually the case when some you know the agent is coming back with some answer usually providing some options or >> uh with some um images with URLs things like that okay >> so the good thing with carousel is that you can have um different tiles for different uh different options and for each of them you can integrate with really interesting UI like an image or a button or you can you can even have actions based on uh on the different buttons you have. So you can really have something really well integrated. >> Is that card based card service? >> Yeah, it is. It is. Yeah. Yeah. So that's a new widget. Um but that's that's really I think that's really interesting to have that in the in again in the in the context of an agent because the agent is providing so many >> um >> so much information in different different form different and that you can illustrate differently. >> Yes. I I think the the next missing tool that we don't have is how to convert markdown to card service. That should >> I would love to have that. [laughter] >> Uh >> well maybe an MCP will do it. >> Uh yeah I don't know would that help like if having a MCP >> that's more like at the SDK level but we could definitely work on something like that. >> Something I would love to see is maybe not something that will generate it live but you know for example when you when you build your agent you specify the um the output uh format so you know what what kind of structure you have. Um so as soon as you have that when you're developing you're supposed to be able to generate you know some sort of prototype of uh visually this can be illustrated like that and then with card and then and then you're good and you can put that in your in your app and >> uh get started pretty quickly or something like that. Doesn't have to be dynamic um at runtime. Yeah, I was exploring with MCP just uh taking a JSON and kind of iterating to, you know, kind of pre-render and let the AI kind of iterate to create the perfect card based upon the prompt. >> Um, still need to work on that because there's some hacks underneath it with, you know, scraping the website for the card tool. But >> yeah, it's an AI tool that would do UI wireframes. Do we like I I like to be able to explain to AI that hey I want a form or something and it produces a wire form even the HTML or JSON that would be amazing. I don't think that exists right now. >> No, I would love you know what I don't know if uh we'll have to cut here but [laughter] uh I would love to see u something with Figma like we design something on Figma. Uh I know they have an MCP as well uh in the box. Okay. Would be awesome. Like if you can uh design your cards in Figma and you have an MCP that does that >> could be really interesting um for all developers. Uh you know if they use Figma with designers and so on that will be a really really good uh uh connection as well. It's not only developers or is showing to the entire company uh oh I'm building this I need the budget or the resources to do it things like that. Yeah, that seems like something we could kind of do is you're taking this kind of this end goal which is this design pattern that you want it to look like that we have the specification for cards and we just need to tie in the pieces kind of I explored this a little bit where the AI can iterate on the JSON until it matches basically with the card output. >> So >> yeah, >> we should show that. >> Yeah, we should. Yeah, uh we should build [laughter] Yeah, let's do it. Uh we're going to do it right after this spotlight. Uh yeah. Uh so I'm going to switch some topics. Uh let's talk about what was some of the most interesting talks that you were were able to attend and anything that excited you kind of made you think about what you're going to work next uh based on what's happening right now? >> Yeah, for me it was just kind of the orchestration layer on top of the AI developer tools. I can get stuck in that and just like try to optimize my processes and to, you know, submit the PR that changes 10 lines. >> I can spend an hour trying to optimize this agent does it this way and I have another agent to review. So I always get excited about kind of how developers are doing that, >> right? >> Yeah. >> Yeah. >> What about people on my side? I think what stand stand up stand out is um is the things people get got excited with you know flows that we are developing right now um >> so on on my side I feel like it's important for people to know that you know it's coming something is coming um and there will be some uh extension opportunities for developers uh or some need to do it so they I think they should >> uh pay attention to it they monitor what's going on >> uh it doesn't mean that the whatever we have today will be exactly what we'll uh launch at the end but >> something is is coming and developers will be part of it in some ways. Um so I think that's that's important and I think people got excited somehow by the use cases and so on. We we got a lot of questions. So uh I think there is some traction you know on flows. Um and the second one is pretty much what Justin said like I I told you I'm still copy pasting but so many of you guys are so much uh you know advanced. moved away from copy pasting. >> Yeah. Well, I I need to play with CLI now. >> So, yeah, one of the things Yeah. Floz was exciting was the I guess the highlight of the our events both here in Paris and in Sunnyvale. Uh but uh yeah, the one major request I had was for custom triggers. Everyone kept coming like hey when's custom triggers coming to flows and hopefully we'll do that. The other one was this is very close and dear to me is uh having a custom action not custom action but a primitive step that just lets you call a appcript function or appcript scripts cuz right now if you want to build call an appcript function you have to build a whole new uh flows add-on. >> So I had flow I had folks from Airbus and whatnot like hey uh I just want to call a custom function but I don't want to build the add-on. So I was like, "Oh yeah, there was my first feature request for this." So >> I think I think that's that makes me think about something we we have for app sheets, right? So it's like it is really famous like it's one of the best feature ever, right? >> Exactly. So all these folks have been using app sheet and they know they can call a custom appcript function and so they come with that mindset to flows too >> and so for them it's like oh I why can't I do this thing that I can do in this other app, right? So yeah, you have to explain it to them. Uh yeah, any other thing that popped out? >> Uh yeah, one of the takeaways I had is everyone loves Appcript and are doing tons of things with it and tons of crazy things with it too. >> Even even on the pre track, they they will like be super advanced in Appcript and prototyping and so on and sometimes just releasing with it. So it's like yeah interesting. >> Yeah. Oh yeah. One of the sentiment about Yeah. Everyone loves Appcript. Um but then or uh one of the many u sentiments I had about appcript it's still stuck in this primitive IDE. >> So like hey why can't I use VS code? I like yeah you can use VS code with clasp and stuff but you're going to miss some of those things. Uh the other one is uh you know we're already calling a lot of AI in our day-to-day and you know we're using apps to call AI but now they want AI support AI code assist within appcript and like hey I want AI to write code for me inside appcript ID and I'm like yeah hopefully that comes someday but those were some of the things that came to me. Yeah. >> Uh >> I think I think we got a talk that u discussed a little bit about the u what's next for appscript. >> Yeah. >> Uh I'm saying that because at some point I guess we'll we'll be publishing uh the sessions uh that we had uh here in Paris and some inl. >> Okay. >> So it could be interesting you know for people to know that um at some point we'll have these sessions uh on demand >> um and you know they can they can they can watch them if they need to learn more about all that. Uh, I think that's important to to note. >> Okay. Uh, so yeah, we're almost at the time. Any last words from you, Justin and Kick, anything you want to add before? >> Uh, no, not really. >> I mean, I would just say check out, you know, the developer MCP server for Google Workspace and the VS Code extension that we have that helps with oath scopes and your editor to kind of kind of guide you on the right path. And those are public repos so they can contribute to that too and make the context more. Yeah. Anything you wanted to add? >> I I I think people are here watching this video from YouTube. I think they should subscribe and >> Oh, yeah. And from there we usually [laughter] do it as as you want, but I think people I mean we we usually publish uh really interesting content on YouTube. So I think that you know that's a really good channel for people to to monitor as well. Yeah. >> Okay. Yeah. Uh, with that, yeah, we'll end our spotlight today. Thank you for attending our talks if you were here in Paris or in Sunnyvale. Hope to see you next time. Thank you. Bye.