Vibe coding has taken off in the past year, and it’s not just a way for developers to play around with side projects. Just as low-code had in the past, vibe coding is another evolution of the movement to bring the power of software development to people without a coding background.
In the latest episode of our podcast, we spoke with Angie Jones, VP of engineering for AI Tools & Enablement at the financial services company Block, about this movement and how it’s happening within Block.
Here is an edited and abridged version of that conversation:
Q: One of the things that people talk about AI is how it can empower everyone to create code and do things they couldn’t do before, when it isn’t now just about software engineers. So how is that changing the way organizations are working, you know, creating now an army of people who can create code that may or may not have any governance behind it, any security, any testing, things like that?
A: I think that engineers are adopting these AI tools, and what we’ve learned is that in order to keep up with them, we have to use AI ourselves. They’re moving really fast, and so within Block, we developed a coding agent by the name of Goose, and our developers use that. But the rest of our employees decided hey, we want to use an agent too. We want help as well, and so that really shaped how we built Goose out so that everyone within the company is able to use this AI agent to do their jobs as well.
Q: I also wanted to talk to you a little bit about this notion that AI agents through MCP are kind of morphing into the enterprise operating system. What do you mean by that, and what does it mean for developers?
A: Everyone wanted their hands on Goose so that they can also be productive. We saw sales and marketing and finance and legal — Everyone was using this tool with their respective MCP servers, which were connections to the apps that they use, whether that be Google Drive, Gmail, Linear, GitHub, or even databases like Snowflake and all of the tools that we use internally. And so we built MCP servers for all of these tools to enable employees to utilize this. And they were using Goose in it, but Goose is still very much a developer tool. And so what we found is that even utilizing it might be a little difficult if you’re not the most tech savvy person. And so what we’ve built now is the next layer, which is an agentic operating system. Think like a dashboard kind of thing, where we have all of these various agents running on people’s dashboards, and they can pull whatever agents they want. Think of an app store, if you will, with all of these various mini agents that people can say, Yes, I would like that to be on my dashboard. Or yes, I want something that maybe gives me updates on the tickets that I’m monitoring, or the top issues that I need to go resolve, or maybe those Slack messages that I haven’t quite gotten to yet. And so they can add these to their dashboard, and all of them are agents behind the scenes that are connected to MCP servers. And so people really love this. You know, it’s a really nice interface where they can have a conversation with the agent, but they don’t necessarily feel they need to pull out a developer tool in order to do so.
Q: I had mentioned earlier about if you just give everybody the keys to the kingdom, so to speak, there can be all kinds of issues. I’m hearing now these days about these opinionated IDPs that kind of create the guardrails that say, well, you can only go this far, or you can only go that far. So what are you seeing in terms of the effectiveness of that kind of an approach?
A: Our security, identity, like all of these folks, are really close to engineering and developing out these tools, because you’re right, the permissions, the access, even like when these agents take an action, are they doing it on their own behalf, or are on yours and is your name associated with that? And so these are definitely things that we’re keeping in mind, as we’re building out these tools, but also informing and educating our employees, like, yes, you’re using these non-deterministic tools to assist you in doing your job, but ultimately, quality and what they produce is still your responsibility. They need that training on how to effectively interact with these agentic systems.
Q: The other thing that people are talking about a lot these days is vibe coding. One of the things that I wanted to ask you about is with AI creating this functional code, where does the human part go into that? Can you prompt, for instance, AI to not just create this code, but have it kind of be the way you would do it if you were writing it?
A: I think the way to solve that is by providing a lot of context, essentially. And so I’m working with our engineering body to help them move beyond just the vibe coding and also outside of the IDE. Within the build systems, when they put up a pull request, we have agents spin up automatically, or even if an issue comes in, they can assign that issue to an agent and have them implement it and put up a pull request and things like that. But in order for it to do that effectively, it definitely needs context, and so we’re doing a lot of training around how to build out rules, files and things that help the agent navigate your code base. We have huge model repos that are 10 to 15 years old, and in these code bases it’s really difficult for an agent to come and grasp the complexity and the nuances of how this code has been developed, and so as a human, we need to provide that information. If we’re going to delegate some of these tasks and ask it to write the code that needs to seamlessly integrate into these code bases, we have to do our part to set it up for success.
