AI Agents and Open Source

Narrator
Hello and welcome to Code[ish], an exploration of the lives of modern developers. Join us as we dive into topics like languages and frameworks, data and event-driven architectures, artificial intelligence and individual and team productivity, tailored to developers and engineering leaders. This episode is part of our Tools and Tips series.

Julián
Hello, hello and welcome to Code[ish], the Heroku podcast. I’m Julián Duque you host, principal developer advocate here at the Heroku, and with me we have a great, great host. I’ve been waiting to record this episode for a while and finally the time has come. Today with me is Rizel Scarlett. She’s the tech lead for open-source developer relations at Block, and we are going to be talking about AI coding agents and open source. Hello Rizel, how are you doing?

Rizel
Hi, Julian. I’m good. How are you doing?

Julián
I’m doing pretty good. Thank you, thank you so much for joining me today here in Code[ish]. We finally relaunched the podcast after like a long hiatus, and we are making sure we are bringing in great voices like yours to this space.

Rizel
Awesome. I’m happy to be here.

Julián
Awesome. So right now, one of the most interesting topics out there for us developers is agentic AI. These new AI agents that not only are helping us to build code, to build solutions, but also helping a lot of companies to automate a lot of work. So, I want you to tell me first a little bit about Block as a business, and second, how this new agentic era is helping Block succeed in their business.

Rizel
Oh, of course I can do that. All right, so a lot of people might not be familiar with the name Block, but you’ve definitely used a lot of our products. Block is the parent company of other companies like Square, Cash App, Title, Afterpay, and a lot more that I probably forgot to name. We’re essentially a bit of a fintech, but we do so many different things and our whole goal is to, like, empower different businesses or people financially, whether they’re mom and pop shops through Square, or whether we’re enabling individuals to be able to easily send money back and forth, or enabling musicians to basically be able to get their money easily through Title. So, those are all of our businesses that we have, and it might seem random, like, why will we get into AI agents, right? Like all of a sudden, from fintechs to AI agents. But the thing is, we went through this period where all of our companies kind of acted as separate companies, but we recently decided to functionalize, or all like work together. And this reorg added a little bit of friction for various engineering teams. Right? So we were like, oh my gosh. Like this engineering team works way differently. But now we’re all expected to work, like, together underneath this one engineering leader. So one of our engineers, one of our machine learning engineers named Bradley Axen, he started developing Goose, which is our AI agent, and he was using that to help our teams with data migrations and all of these other things. Just kind of automating tasks for us. And we found it to be super helpful and it started getting bigger and bigger. He started adding more to it. We like made a team around and we’re like, oh my God, this is good. We need to open source it, right? And then we started working on an idea. We were like, oh my gosh, what if we can have our AI agent be connected to different sources? We connected with Anthropic who had a similar idea, and they’re like, we’re actually working on model context protocol right now. So we connected with them. We contributed to the spec. And that’s how Goose came to be, where it was something, and it continues to be something that helps automate tasks within the engineering teams. Even people who are not engineers use Goose. Like it’s used across the company. HR, executive assistants… And we also have it open source because we’re like, this is really cool, and we want other people to be able to benefit from this.

Julián
So that’s beautiful. So even started as an internal agent and internal tool. And then you make it public through open source.

Rizel
Yes. That’s how it started.

Julián
Well that’s a that’s the wonder of, of open source. I was checking more about Block and one thing that they really like is that open source is at the core of the company. And I bet not only Goose is the product that you have open source, but you have like other projects there. Which ones are the ones that you maybe use more, or do you like more, about the open source offerings that Block has?

Rizel
Okay, I mean, in terms of I mean, obviously Goose because that’s where the industry is at right now. It’s all about AI, but another one that I like is Hermit. Goose actually uses Hermit under the hood. And this is like a type of package manager. The rest I’ll be honest, I don’t get to use them in my day to day because I, even though I’m the tech lead of open source devrel, our company focused and goals are all on Goose.

Julián
Oh, nice.

Rizel
Yeah.

Julián
So betting 100% in AI right now.

Rizel
Yeah. As other companies are doing as well.

Julián
Of course! And it’s very interesting because all that real usage of these LLMs and these agents is helping the whole ecosystem evolve. Initially, for example, with MCP code release. Sure, the concept behind was very good and powerful. And you can see, okay, I’m giving access to these LLMs to execute tools or have like expanded context, but initially there were a lot of gaps. Okay, we have not solved the authentication problem right now. Maybe the protocol we are using for communication is not the best. But now you see that the spec has evolved thanks to this enterprise usage. And as you mentioned, you also contributed, like the company contributed, help evolve in that aspect. So the more companies are using these technologies and of course, the more we can see open source being out there, I think the evolution is going to be better and we all are going to benefit from it.

Rizel
Oh yeah, I completely agree. I think that’s always my thoughts when people are like, what’s the big deal about MCP or model context protocol though? Like it’s nothing game changing. And yeah, it’s not like this really complex protocol, but it just provides this ease of use. And since so many people, whether from community or just like you said, enterprise companies have bought into this, it makes it easier for everyone to kind of go around this ecosystem or community and continue to build up from there, rather than having all these separate solutions. Having that like, those decentralized solutions doesn’t really help automate everyone’s workflow, because you have to do it in all these separate, different ways.

Julián
So I have not used Goose, and I’m coming here as a person that doesn’t know about it, and I want to learn more. So tell me a little bit more about Goose. Is this a coding agent, or is this an automation agent, or what is the main use case for let’s say, a person like me to get to use Goose as a tool?

Rizel
Yeah. I mean, yeah, its specialty is definitely automating engineering tasks. So it helps you with coding. But as I mentioned before, it’s really kind of all purpose, at least from my point of view, because people within my company that are not engineers use it day to day. So you can use it if you wanted to, like help code an app. And that’s how I mainly use it. But we have other business areas using it to like figure out, okay, I missed, a couple weeks or I went on vacation. Right? And I don’t know what I missed from Slack and Google Docs and all that. So they’re literally just using Goose, connecting it to like a Google Drive MCP server or a Slack MCP server and saying like, hey, what did I miss? Or hey, like, what are some things I should highlight on my performance review? Or even on the analytics side? A lot of people tend to use that to help collect and analyze data for them. But yeah, mainly I use it for coding and I think that’s like the main use case, but there’s so many others.

Julián
Interesting. I have seen, for example, a lot of people, and especially in our teams, that they use Cursor, for example, to solve non-coding tasks. But they still, for me…sure, you can do it. You have an agent that can execute MCPs, but the interface might not be the right one to be doing, for example, that “go to my Google drive, fetch documents and bring me an analysis.” It can do it, but it is not the right tool for the job. I see the IDE for coding, so what is the main difference? For example, you mentioned that the you use it for coding. You use it like a standalone tool, or it is integrated within an IDE. How that experience looks like?

Rizel
Yeah, I love that you pulled on this thread for me. Yeah. So, you have two different interfaces. First of all, it’s a standalone tool that’s local to your computer. So it’s not in the web browser. It’s not connected to an IDE. You can either use it as a desktop app. This is like an Electron app that you can like download and you can it has like a chat interface. So you can just like chat with it and be like, oh, can you like go to my Google Docs or can you access my local system and like organize my folders or whatever it is you want it? And then you have the CLI interface. And this is where developers would probably be more, they might lean more towards using this for coding because there’s a CLI interface and they’re already in the CLI, but you have both options for non-developers and developers. And like the way we look at it, integrating with MCP servers is that we, we don’t even refer to them as MCP servers. We call them extensions because they’re extending the capabilities of Goose for developers and non-developers.

Julián
Yeah, MCP is just the protocol behind. I agree, they are just extensions or maybe integrations. I will use a different name. That’s great that you are also relying on, just open source, open standard technology for that integration, for that extension. And one thing with, for example, Goose, you can connect any specific MCP or any specific extension, or you have a dedicated centralized registry or repository of buried extensions that you have. Let’s say I want to use Goose today with the Heroku MCP server. We have an MCP server that lets you interact with the Heroku resources like deploy an application, provision a database, fetch logs, etc. Can I do that today? Or I just can use the MCPs that are approved by Goose.

Rizel
Oh you can do that today. You can literally connect any MCP server as an extension to Goose. However, we do have a list of extensions that we vetted. And then we made, like, tutorials for people so they can easily just click and go. Like, there’s deep links. You could click it, it’ll automatically install. So you don’t have to like create a Json file or add anything to like you configuration. Just making it easy. But since you mentioned the Heroku one, we got to work together to add Heroku to the list.

Julián
And it was not the plan just to do that type of collab, but sure, why not? We have two, the local one, which is the standard input and output that supports most of the functionality from our CLI. Basically, the approach we did is we have the Heroku CLI that you use to interact with the Heroku. And we added the one command, which is the Heroku MCP, that starts the MCP server and exposes the CLI tools to the client that uses the MCP. But we also have a remote version of the MCP, which we have unauthenticated. The endpoint with a limited set of tools. So we have the two. Does Goose support both local and remote, or just local?

Rizel
Yeah, we support local, remote, streamable, HTTP, SSE. All of them. Yeah.

Julián
Awesome. Awesome. I need to spend some time then to give it a try. And why not, I will let you know. Okay, this is how I connected to Heroku, this is what I was able to do. And maybe we can, we can do something together for sure. I would love it.

Rizel
Yeah. Me too.

Julián
Yes. And what other, for example, use cases that Block customers have found by using Goose? Are the customers also using Goose, or is this is mostly a tool that was created for engineers and developers out there. Is there a specific use case for customers?

Rizel
Interesting question. Okay, so my focus is more on, like, the open source and what developers are doing with Goose. And then there’s, like, my counterpart who’s focused on like, internal, and then what our customers are doing. But I know one thing we do have is a Square MCP server, right? And Square is like, where a lot of our customers who run mom and pop shops use it to like, figure out like, oh, what is in my restaurant? Like, what’s my menu? How are people paying? Like, it’s a point of service or like a point of cach type of system for people. So, I think a lot of our customers tend to use Goose through the Square MCP, or integrated with the Square MCP.

Julián
That’s a beautiful because that’s an interface that it’s serving all of the different audiences. And I think that’s the key part. And one of the most interesting things about these new AI tools is the ease of use. So a chat or voice conversation, I’ve seen recently more voice-enabled chats for AI applications. And I think it’s a very natural interface.

Rizel
Yeah.

Julián
You don’t need to be an expert programmer or a person that is well versed in using a computer or be an expert on typing. You just, like, click on the microphone and start having a conversation. And usually these agents are going to respond with like a natural language or a natural voice. Do you see that maybe, are going to be plans to add voice to Goose in the future, or already has that feature?

Rizel
We have voice dictation.

Julián
Oh my God. See?

Rizel
Yeah. You know, it’s crazy. Jack Dorsey, our CEO, he actually added it. One of the cool things about Goose, because it’s an AI agent that can help you code, a lot of times we use Goose to help us add more things to Goose. Like, I’ve done it myself. It’s been a really great enabler for me, and we watched our CEO do it. He sometimes like hops on a Google Meet and we just all watch him essentially, like vibe code. And there was one time where he decided he’s going to vibe code, adding voice dictation so you can speak to Goose. And I seen a lot of, in addition to like you just a non developers use it, I’ve seen engineers use it and I was like really like, this is not my workflow. It doesn’t really work as well for me. But I was really surprised that a lot of engineers said like, it’s much easier for them to talk out what they want to do, their strategy, or just think about like, oh, how is this built in architect things when they talk it out loud rather than, than writing it?

Julián
This is a whole new version of rubber ducking. Instead of you talking to like a rubber duck that is not going to reply to you, now the agent is giving you answers and then giving you tips.

Rizel
Right? Isn’t that crazy?

Julián
It is awesome, I think. This keeps getting better and better. These these tools. Speaking about Jack, I saw an article by Angie which is the VP of DevRel at Block. talking about like best practices around vibe coding while watching Jack doing it. So we are going to make sure we share that blogpost within the show notes, because I found this fascinating. And definitely by looking at other people, their workflow, how they interact with these tools, you you always learn something.

Rizel
Yes. Yeah, you learn a lot. You learn so much. Yeah. Our team loves watching Jack like vibe code because we’re like, wait a minute, we didn’t realize you could do it that way. And then I also like run a live stream called The Great Goose Off. And I basically challenge people to, like, create ridiculous things with Goose, you know, like, do like a face off and like this person against this person. I’ve learned so much, where like somebody said, like, if you’re using a GPT model and you’re using Goose, if you tell it to think harder, it’ll go through like a deeper thought process. And I’m like, oh, I had no clue that that’s how it worked.

Julián
Oh, so you can select also the inference model of your choice while using Goose?

Rizel
Yes. Oh wow, I really didn’t sell Goose that well in the beginning. Yeah. It’s bring your own model. So whatever like provider that you’re used to using, whether it’s Anthropic, some people even have been using GitHub Copilot as a provider. That was like a really requested one. I didn’t realize that GitHub Copilot does that. But yeah, all these different providers, OpenRouter or Databricks, because it’s open source people have just been even adding ones that we weren’t aware of. And you can basically, as long as you have the API key, plug it in and use it with Goose.

Julián
Now this is giving me more ideas. And I love open source. So we here at Heroku, we have also managed inference services. So the models are hosted in our infrastructure. So for our customers they just provision an AI model with just one click or one CLI command. We support Claude, we support GPT, open source. We support Nova from Amazon. We support embeddings, models, emails, generation. So it will be interesting. Also it is OpenAI API compatible?

Rizel
Yeah I think you could add it.

Julián
So he’s just like writing a connector. I did recently the connectors for LangChain both in TypeScript and Python for the Vercel AI SDK. And again to connect our AI services. So that has been kind of my side projects. Not part of my job, but I just enjoy, like, doing open source and connecting with these different frameworks out there.

Rizel
Yeah, you got to add it to Goose, then because we have, I think, a similar counterpart of what you’re saying. We have Docker Model Runner as a provider. That’s like something you could add Heroku in. Okay, so I didn’t say all the cool things about Goose, maybe things that make it stand out.

Julián
Tell me other cool things!

Rizel
Okay, first of all, we have this thing called recipes that kind of help us stand out from other AI agents, because a lot of times it’s basically a reusable AI workflow for folks, right? So a lot of times when you use an AI tool, whether it’s an agent or like ChatGPT, you tell your friend, oh, just give it this prompt. And obviously AI is non-deterministic, so it’s going to give out a different result because you don’t have the same configurations and everything. So what recipes do is it packages like the settings you had, the configurations you had, whatever MCP servers and extensions you were using. And you can essentially send that recipe as a link to people or reuse it just for yourself, and always get back the exact results that you need. So I think that’s something that’s really interesting and helps us to stand out. Like I’ve seen people use it for teams like to onboard engineers, or they just have this specific workflow that really works well for them. I know, Angie it to help her like keep track of our work. So just like do a recipe and say like, hey, what’s my team working on right now? And it’ll pull up all those stats where she doesn’t have to reprompt it again and figure out which extension she needed to have and able to pull all of that information.

Julián
That’s a great DX / UX approach, and it is one of the things I mean, for example, with these other coding agents, I have, let’s say, my set of rules and different agent definitions with what I want to use. But sometimes I find the difficult, one, sharing those with other people, or having a repository of different rules that I want to enable and disable. There is not like a native way of doing it in these tools I am using. Right now is super like a very simple approach. And that approach seems pretty smart. Like I package it and then I share it and automatically load and my agent is pretty much what I’m looking for.

Rizel
Yeah! Try Goose! I feel like I’m a bit of a salesperson. But yeah, I think that that will work for your solutions. You can even have, like, sub recipes which is, like, okay, maybe your recipe is taking a little too long or the task is looking too complex. You can piece it up into different parts and be like, okay, I have this sub recipe. So my recipe will run multiple sub recipes in different steps. It kind of reminds me of a GitHub action, because you could write it as a Yaml file and have like those steps in a particular job that it goes through. Or you can just have it, as I mentioned earlier, as a deep link.

Julián
Nice.

Rizel
Yeah.

Julián
And it is not that you sound like a salesperson. It is that we, developer advocates, we need to be believers and passionate about the product we work with. And I can tell that you like what you do, and I can tell it in your voice, the passion that you are pretty much projecting while speaking about these tools. Which is pretty much why we do what we do.

Rizel
Exactly. I wouldn’t be working with this if I didn’t think it was really awesome.

Julián
Exactly. So do you have any tips for folks that want to start using this type of coding agents, or to start like, using these tools to solve specific problems? Do you have, like, a workflow to share or like, specific advice for folks so they can be better at using these tools?

Rizel
Yeah, I would say start small is my first thing. I know some people, if you’ve never touched a coding agent before, it can feel overwhelming. Or you might be asking it to do a bit too much and then be like, oh my gosh, this doesn’t work. But start small with asking it certain things and then like get bigger. Like once you get comfortable, do bigger and bigger requests and then my tips may lean more on the side of like, being responsible for like coding while using a coding agent. Whether you do like vibe coding, quote unquote, or like just AI assisted coding, I think sometimes people get a little too comfortable with vibe coding, and you just trust what the AI tool generated. But just like you would check the work of your coworker or like whoever else you’re working with or your own work, you should do that the same way for an AI agent. And with platforms like Goose, there’s a lot of things built in to make sure that the output is refined, like Goose has Goose ignore files, which is kind of like a Git ignore. But basically you’re saying like, “Goose, don’t touch this particular file. I want to protect it and save it.” You can, I think you mentioned like having hints or rules? Like Goose has goose hints, similar to like Cursor rules. Also try to make a plan ahead of time before you just jump into like, hey, I want you to build this application. It may go in whatever direction it interprets. So try to like work with Goose to make a plan. There’s a lot of ways you can actually do that. We have like a plan command that helps you to align your agent before any code is touched and like use version control too. That’s like my top tip.

Julián
Oh yeah, that’s super important! I have learned my lesson the hard way by not using version control in time.I always use it, but sometimes I get too excited. Okay, let’s continue iterating and then everything is broken and I had to ask the agent to roll back and it doesn’t do it. It gets a little bit painful.

Rizel
Yeah. That’s like I know this as a rule or a good guidance. And I tell everybody, but sometimes I forget because the same reason you said like I get that excited, I’m like, oh my gosh, that worked. I’m like okay, now do this. And like it’ll be something small. We’ll be like, change it, blue. And it changes the whole app. And now like I can’t command Z for every file or like tell Goose or whatever, can you change it back and it only changes that part of it. So yeah, version control is key. Actually what I do is I use the either I tab Goose and use the GitHub CLI or the GitHub MCP server, and I write a Goose hint that says like after every change, do a commit.

Julián
Okay.

Rizel
That way if I forget I could just easily roll back. Yeah, there’s so many tips.

Julián
Really solid tip. I want to add something, complimenting what you started talking about. These tools are opening a lot of doors for people that usually were not like developers. Sure, they have the idea, but maybe they didn’t have, like the full skills or the confidence to start like sitting and writing code all day. But at the end, you are still coding and the best practices, all of these fundamentals, are still valid, even though we have amazing tools that can build a full blown software as a service from scratch, without the person not writing a single line of code, having those fundamentals are pretty, pretty important. So one advice that I have for people that want to start with this is don’t ignore those fundamentals. Keep learning. Keep paying attention to that. This is not going to replace us developers, like a lot of people is saying. I don’t see that in the future. It’s enabling us. And it’s pretty much making us focus on different things, on improving the productivity. But the fundamentals are super important, so don’t ignore those. And this is more a message for our audience. Don’t ignore those because you are going to need them for sure. You are definitely going to need them.

Rizel
Yeah, yeah I think that’s a lot of people’s worry is like, oh my gosh. Like some people don’t like the idea of vibe coding or AI assisted coding. That they’re like, now people are going to rely too hard on the AI agent and they’re not going to continue learning. But I believe that you can, you can use AI tools to help you learn. If you’re like reviewing the code and, and just getting exposed to new tools. And I’ve even seen people there’s a contributor who made an MCP server called Goose mentor mode or something like that, and he uses it to teach his students about code. So it’ll like ask them questions and have them reflect on like why they’re doing or like why they’ve implemented certain things.

Julián
That’s beautiful. Yeah. We need to keep using all of these, these technologies, getting advantage of them, through open source content creation and these type of channels, we are in right now sharing with the rest of people so they can also benefit from this. We kind of gatekeep because this is, this is something that I think it can benefit a lot of people, even though sadly some jobs are being replaced with these new workforce replacement. It is opening a lot of new jobs as well.

Rizel
Yeah.

Julián
So it is…embrace change. This has always happened in our history. Embrace change and learn this so you you can master it.

Rizel
Yeah I look at it like the horse and buggy versus the car. Maybe when cars came out I imagine some people were like this is a threat to my job. And of course in some way it was because they’re not looking for people that operate horse and buggy where they’re looking for drivers. But like you said, we need to embrace the change of AI, adapt, learn some of those skills. That way you’re not left as just knowing how to do a horse and buggy and nobody’s using that anymore. Learn how to figure out how it can make you more productive at your job, and be more of assistance, rather than just replacing you, is. That’s how I like to think of it.

Julián
Exactly. That’s how I that’s how I see it as well and how, when people ask me about it, it’s kind of like anxious about the future. I told them this is going to be beneficial if we take advantage of it in the right way.

Rizel
Yeah. Agreed.

Julián
Well, Rizel, thank you. Thank you so much for all of this wonderful conversation, for telling us about Goose and all the open source wonderful things that your team at Block are doing. I would like to invite people to go ahead, use these tools, explore them, try to solve those problems that you always do on your work, on your day to day life, and try to be a little bit more creative by relying on AI. I don’t be afraid of embracing these, just approach it and learn because this is definitely going to open you a lot of doors. Rizel, do you have any last words for our audience today?

Rizel
My last words I will say is I just want to emphasize the idea of embracing that change and just figuring out how it can work for you. This is a huge time in the industry. It may seem scary, but this is just like when people came out with the internet and everyone was saying everyone should be using the internet. Don’t let yourself get left behind. And also make sure you’re a voice in the AI industry or in the direction of AI, because it could take a turn that we wouldn’t like. If you don’t let your voice be heard and contribute through open source to basically guide that direction.

Julián
Yes, 100%. Well, thank you again and thank you. Thank you so much to all of our audience for listening to this Code[ish] episode, and I’m looking forward to see you on a next one. Bye bye.

Narrator
Thanks for joining us for this episode of the Code[ish] podcast. Code[ish] is produced by Heroku, the easiest way to deploy, manage, and scale your applications in the cloud. If you’d like to learn more about Code[ish] or any of Heroku’s podcasts, please visit Heroku.com/podcasts.