At a company-wide hackathon this month, developers at finance firm Block built a dizzying number of prototype tools including a database debugger, a program for identifying duplicated code, and an app that automates Bitcoin support.
The sudden productivity boost was driven by Goose, an artificial intelligence agent developed by Block several months ago that can help with coding and other work like knocking together data visualizations or mocking up new product features.
“We’ve always had really strong hack weeks, but this one was at another level,” says Jackie Brosamer, who leads the AI and data platform at Block. “We have tens of ideas that we’re looking to bring to production.”
Goose helped developers at Block to develop a new agent-to-agent communication server at the hackathon. The company says Goose has changed the way it works, not only helping automate code generation but also allowing non-engineers to dabble in coding or prototype for new apps or features.
I first spoke to Block several months ago, when Goose was a little less cooked than it is now. Developers at the company admitted that the agent increased their output but at the time also sometimes made mistakes like deleting the odd file (this can still happen sometimes). They ran the system on machines where any changes could easily be rolled back.
Agents are starting to change the way many developers and companies operate as AI models get better at managing code, using computers, and wielding tools. Over the past week, Google, Microsoft, and OpenAI have all touted agentic coding tools. Block uses Anthropic’s Claude model by default, which is particularly good at coding and tool use.
Block CEO Jack Dorsey and the company’s CTO, Dhanji Prasanna, concluded that agents would shake up their industry in fall of 2024, when improved AI models triggered a significant leap in the performance of many agents. Dorsey and Prasanna decided that Block should build its own agent and that engineers and other staff should dive headfirst into using it.
Block’s Goose is available as open source (the name, in case you didn’t guess, was inspired by Maverick’s friend in the movie Top Gun).
Goose can be powered by a range of different AI models and will run commands and access files and folders on a computer. Goose can also tap into a growing number of online tools, like cloud storage platforms or online databases, thanks to the Model Context Protocol scheme for agent communications developed by Anthropic.
I used the latest version of Goose to knock together a few simple games and a basic visualization. It does a nice job of handling tedious things like ensuring the right version of Python is available and installing packages. Other tools I’ve tried seem as capable, but the Goose interface is particularly easy and intuitive, and it seems likely to become more powerful as it gains access to other tools and services.
Brad Axen, an engineering lead at Block using Goose, told me recently that Goose has proven especially useful for helping engineers wrap their head around an unfamiliar code base. “One of the things that works best is if you say ‘Hey, I don’t know how this works yet, could you please go find everything and give me a summary,’” he says.
Axen said back then that using an agent seemed to require a different mindset where you didn’t expect the tool to be perfect and were prepared to roll back any missteps.
Block says that some developers were initially resistant to using Goose, reflecting skepticism I’ve heard from other coders who don’t trust AI to code well or are reluctant to hand over their own agency. But the company has developed protocols for managing code produced using Goose, for example having humans review everything made at the hackathon for bugs before it is cleared for release. As the complexity of these agentic systems grows, managing stability and security may become a bigger challenge.
Axen says he’s most excited to see how the server built during the hackathon can change the way other agents use Block’s technology. “It’s a weird way of thinking about computers working together,” he says. “But it’s exciting.”
What do you think of Goose and the way agents are changing work? Send an email to [email protected] to let me know.
