At a big company hackathon this month, developers at finance firm Block built a ton of prototype tools, including a database debugger, a program for finding duplicated code, and an app that automates Bitcoin support. The sudden productivity boost was all thanks to Goose, an artificial intelligence agent developed by Block several months ago. Goose can help with coding, data visualizations, and mocking up new product features.
Jackie Brosamer, who leads the AI and data platform at Block, said, “We’ve always had really strong hack weeks, but this one was on another level.” The company has a bunch of ideas they want to bring to production now. Goose helped the developers at Block create a new agent-to-agent communication server during the hackathon. It’s changed the way they work, automating code generation and allowing non-engineers to try coding or prototype new apps or features.
When I first talked to Block a few months ago, Goose wasn’t as advanced as it is now. The developers mentioned that the agent increased their output but sometimes made mistakes, like deleting files (which still happens occasionally). They were cautious, running the system on machines where changes could be easily rolled back. Agents are starting to change how developers and companies operate as AI models improve in managing code and tools. Google, Microsoft, and OpenAI have all been talking about agentic coding tools recently.
Block uses Anthropic’s Claude model by default, which is great at coding and tool use. Block’s CEO Jack Dorsey and CTO Dhanji Prasanna realized agents would shake up the industry in fall of 2024 when better AI models led to a big leap in agent performance. They decided Block should build its own agent, with everyone at the company diving into using it. Goose is available as open source and can be powered by various AI models, running commands and accessing files and folders on a computer.
Brad Axen, an engineering lead at Block using Goose, said the agent has been super useful for helping engineers understand unfamiliar code bases. “One of the best things is if you say, ‘Hey, I don’t know how this works yet, could you find everything and give me a summary,’” he explained. Some developers were hesitant to use Goose at first, reflecting the skepticism of other coders who aren’t sure about AI coding well or giving up control. Block has protocols in place for managing code produced using Goose, like having humans review everything for bugs before release.
Axen is excited to see how the server built during the hackathon will change how other agents use Block’s technology. “It’s a weird way of thinking about computers working together,” he said. “But it’s exciting.” What do you think of Goose and the way agents are changing work? Send an email to hello@wired.com to let me know.
