Block

About two and a half years ago, CEO Jack Dorsey sensed the industry was shifting and created a weekly executive meeting with 40 of the company's top leaders to discuss emerging trends. Donjie Prasanna, then a part-time senior engineer helping one of the engineering teams while raising a young child, observed that despite discussing many important topics, no one was paying serious attention to AI. He wrote a letter to Dorsey arguing that Block needed to become an AI-native company, not just adopt AI tools. After Dorsey visited Sydney and spent two days walking and talking with Donjie, he offered him the CTO role. "It's like, be careful what you're good at," Donjie recalled.

Goose started as an internal desktop tool and AI agent built on top of large language models like Claude and OpenAI's GPT. What made it different from existing AI chatbots was the Model Context Protocol (MCP)—formalized wrappers that exposed enterprise tools like Salesforce, Snowflake, and SQL databases to the LLM. "Until that point, the LLMs were not really able to do much other than chat. But goose gives these brains, arms and legs to go out and act in our digital world," Donjie explained. The tool could orchestrate complex workflows: a user could ask it to build a marketing report, and Goose would write SQL queries, analyze data in Python, generate JavaScript charts, and email the final PDF—all without human intervention. One engineer even had Goose watch his screen 24/7, anticipate his needs from Slack conversations, and open pull requests for features before he'd finished discussing them.

Block didn't need to find external customers—they had thousands within their own organization. But adoption required cultural change. Donjie's first move was reorganizing Block from a GM (general manager) structure, where each business unit (Square, Cash App, Afterpay) operated independently with separate engineering teams, to a functional structure where all engineers reported to a single head of engineering. This was painful but essential: "Conway's law can be really, really powerful... you ship your org structure." With functional teams aligned around shared tools and language, Goose became the natural hub for AI automation. Engineers, support teams, legal teams, and risk management teams all began using it. Non-technical people building internal software tools saw the most dramatic gains—weeks of work compressed into hours.

Block measured success through multiple lenses. Engineering teams using Goose daily reported saving 8-10 hours per week. Across the company, data scientists calculated that AI tools were saving roughly 25% of manual hours. But Donjie was candid about limitations: AI excelled at routine tasks, automating UI tests, and helping non-technical teams build tools, but still underperformed humans on deep architectural decisions, race conditions, and design judgment. The biggest productivity win wasn't AI at all—it was the organizational restructuring that reduced silos and aligned technical strategy across the company.

Goose also showed unexpected use cases. Gosling, a mobile version for Android automation, transformed QA by replacing armies of contractors clicking through screens. One risk management team built an entire self-service system for enterprise risk—work that would have been queued for months on a product roadmap.

Instead of holding Goose as proprietary technology, Block open-sourced it completely. "We believe in the power of open source," Donjie said. The decision paid off: Databricks, mid-tier tech companies, and even competitors began adopting Goose. Donjie was pushing the boundaries further, experimenting with agents that work for hours autonomously overnight, anticipating what engineers want to build before they articulate it. He was rewriting features from scratch daily, embracing a world where AI makes continuous delivery and zero-based design possible. For hiring, Block shifted from viewing engineers as commodities to seeking people with a learning mindset—not necessarily AI experts, but those eager to master these tools and rethink how software gets built. The future Donjie envisioned wasn't one engineer with four jobs, but teams where humans set the goals and guard the taste while AI handles the execution.