Block 怎样变成 AI 原生大公司：Goose、职能制，以及把代码整段扔掉

Dhanji Prasanna, Block's Chief Technology Officer, says the company's AI push began about two and a half years ago. Jack Dorsey was meeting weekly with roughly 40 top executives, and Prasanna noticed that nobody in the room was paying attention to AI. He wrote what Lenny Rachitsky calls his "AI manifesto", although he says there was little to it beyond arguing that Block should act centrally and become AI-native.

At the time Prasanna was not CTO. He was working part-time after having a child and helping one engineering team. Dorsey flew to Sydney and spent two days walking around the city with him, then offered him the job.

His first goal was to make Block think like a technology company again. He says the company had suffered some identity drift, calling itself fintech or financial services when it had started as Square in the mold of Google or Facebook. He started programs that brought top individual contributors together, ran eight or nine small special projects with two to five engineers each, and revived the company-wide hack week.

Block had run Square, Cash App and later Afterpay and TIDAL as what Prasanna calls a GM structure. Each acted like an independent company with its own CEO, engineering practice and design team, and only legal and some platforms were shared. That suited a mature company, he says, but it made it hard to focus deeply on industry-changing technology.

Block now has a single engineering team and a single design team, each with one head. Prasanna compares this to what Steve Jobs did when he returned to Apple. He says it was not a copied playbook but something the team discovered while working out how to bring engineering and design first.

The change was painful, and Prasanna says not everyone was on board. He cites Conway's Law: you ship your org structure. Before the reorg the silos were not aligned on technical strategy, and now people share the same tools and policies, a seniority level means the same thing across the company, and engineers can move to areas of need.

Prasanna says the most AI-native teams now build largely without hand-writing code. Teams with large legacy code bases feel the change less, but they still see background tools that run around the clock or in CI. These tools scan for vulnerabilities and try to patch bugs filed in tickets, so engineers find draft fixes in the morning.

Block's top priority is what it calls "automate Block". Prasanna says very AI-forward engineering teams using Goose daily report saving about eight to 10 hours a week. The figure is self-reported, and Block checks it against PRs, feature throughput and a formula built by data scientists, which suggests roughly 20-25% of manual hours saved across the company.

In his view, the value of AI changes every day, so companies should be adaptable and expand to where it works best. Engineering gains vary: greenfield code bases and apps for new platforms see aggressive gains, while complex existing code bases lag. Rachitsky adds that this is the worst the tools will ever be.

The most surprising use is non-technical teams building tools for themselves. Prasanna describes an enterprise risk team that built a self-service system, compressing weeks of work into hours. He also mentions Gosling, a mobile counterpart to Goose that drives Android through the accessibility API and automates UI tests that once needed an army of contractors. AI still underperforms humans on architecture, race conditions and orchestration.

Goose is a general-purpose AI agent that you install on your computer and talk to like a chatbot. It can organize photos or write software. According to Prasanna, it works through the Model Context Protocol, which Anthropic introduced and Block contributed to early on. MCP is a set of formalized wrappers around existing tools such as Salesforce, Snowflake or SQL that exposes them to an LLM.

Until then, he says, LLMs could mostly chat, and Goose gives these brains "arms and legs" to act in the digital world. It is model-agnostic, working with Claude, OpenAI or open-source models via Ollama. His example is a marketing report. Goose pulls data from Snowflake, Tableau or Looker, writes SQL and Python, builds charts, produces a PDF or Google Doc and emails it.

Block lets employees use any tools they want, but Goose is the best integrated because an MCP is easy to write, and Goose can write its own. Teams no longer wait for a vendor to add AI to an issue tracker or calendar. Prasanna says many companies, including competitors and partners, use Goose, and he names Databricks.

Block gave Goose away because of its belief in open source, he says. As a tech company it depends on Linux, Java and MySQL, so it feels an imperative to give back and to build things that outlast Block. The name is a Top Gun reference, since the engineer who came up with it looks like the character.

Prasanna describes a colleague on the core Goose team who built a system that watches his screen and listens to him. When he discusses a feature on Slack, a few hours later Goose may have tried to build it and opened a PR. It also reschedules meetings when he says he has a conflict.

Prasanna calls it mostly a side experiment, and says you need the stomach to tie an agent that closely into your tools. Still, it shows what is possible and could cut a great deal of busy work. He adds that Block was waiting for calendar vendors to build such features and no longer needs to.

Prasanna says vibe coding is exciting but limiting because it is ping-pong: you prompt, wait a few minutes and nudge a half-baked result. Goose's median session is five minutes and the average seven, and Block is experimenting with next-version sessions that run for hours. He objects that LLMs sit idle overnight and at weekends and says they should build in anticipation of what humans want.

Because humans had limited bandwidth, they had to pick one path to try. He would rather describe several experiments in detail, sleep, and delete five or six of the results in the morning. He writes code daily and now regularly throws away large amounts of it. He asks what the world would look like if every release deleted the app and rebuilt it from scratch.

Rachitsky notes the standard advice never to rewrite, because you lose years of small fixes. Prasanna agrees the trick is making the AI respect those incremental improvements as part of the specification. When he asked a new Goose build to propose and implement its own improvements overnight, he says it succeeded on about 60% of well-described items and struggled on the other 40%.

Rachitsky jokes about giving an AI the goal of driving revenue and having it fire everyone. Prasanna says he does not expect that and thinks human taste is needed to anchor AI so it does not go off script. He says Block's design teams are pushing this, and that taste is a differentiator beyond the era of AI slop.

AI is also weak at judgment across a whole portfolio. When one team wants to buy a vendor tool and another wants to build it with Goose in half the time, a human may ask whether the process is needed at all. He tells teams, especially InfoSec, to question base assumptions, since not building something shrinks the surface area to secure.

On whether AI will replace SaaS, he warns of drifting from your core purpose, which at Block is economic empowerment. Replacing a vendor tool to save money is probably not worth the lost mental bandwidth and technical focus. Rachitsky adds that maintenance is the hidden cost, and Prasanna recalls building Cash Card's core in about a week but spending a long time on edge cases like double tips and gas-station billing.

Prasanna says AI has not yet changed how many people it takes to build something the size of Cash App. What changed headcount planning was the move to a functional structure. Under the GM model, engineers were treated as a commodity and more were added to build more features, while now the company leans on common platforms and experts from across Block.

In hiring, he looks for people who embrace AI tools and want to learn, such as an eager new graduate or a veteran who has figured them out. He is not hunting for AI practitioners as such, and he echoes Dorsey's wish for a learning-first company. Block is starting to let candidates use vibe-coding tools in interviews, but he says critical thinking and deep technical understanding still matter more than AI fluency.

On who benefits most, Prasanna says both the most senior and the most junior engineers adopt the tools eagerly. Seniors are relieved to offload tasks they have done many times, and juniors blitz through work. The real surprise is non-technical staff, who he says are blurring the lines between legal, risk, engineering and design.

Rachitsky notes that fewer one-off requests to engineering should be a productivity gain. Prasanna agrees but compares it to building a bigger highway: more people building software creates more software to coordinate and a bigger appetite for features. He also says Block's test selection cut about 50% of test runs, but offloading tests to the cloud or deleting obsolete tests could save two to three times that.

Prasanna says adoption at Block was driven by leaders using the tools. Dorsey, Prasanna and the executive team use Goose and other AI tools daily. He says this teaches you more about changing your organization's workflow than think pieces on LinkedIn or the Harvard Business Review.

Rachitsky agrees and suggests picking a concrete problem, for example building a Lovable app to pull images out of a Google Doc. Prasanna shares his own: he needed to gather receipts for his son's therapies, spread across screenshots and PDFs, to send to his wife for an insurance claim. Goose tried several approaches, then used AppleScript to put them all in one Apple Note.

Goose is available to ordinary users on Mac, Windows and Linux, as an Electron app with a command-line option. Prasanna says it sits between a classic assistant and focused tools like Cursor or Claude Code, because it is an extensible platform built around MCPs. Users pay only for model tokens.

Looking back on two years as CTO, Prasanna says he would tell himself two things. First, appreciate how hard it is to change outcomes without changing the relationships between people. Second, you only hear about things when they go wrong, so you must make time to step back and check you are focusing on the right problems.

His most counterintuitive lesson is that code quality and product success are unrelated. He recalls the angst at Google over YouTube's code base, with videos reportedly stored as blobs in MySQL on a slow Python stack. Google Video supported more formats, higher resolution and hour-long uploads, yet YouTube blew away its competition, and he says engineers keep wanting to refactor code that could be thrown away tomorrow.

On early Cash App, where he was head of engineering from about 10 engineers to 200 plus, things looked chaotic from outside. His philosophy was to give brilliant engineers freedom, with a safe foundation under the risky parts. He calls it controlled chaos, with limits on liability and money.

His leadership advice is to start small, quoting the line about not boiling the ocean to make a cup of tea. Goose began as one engineer's side project, Cash App began as a hack week idea, and Block's Bitcoin product came from a hack week with Dorsey and another engineer. By contrast, he says Google Wave had 70 to 80 engineers before it had users outside Google.

In Rachitsky's Fail Corner segment, Prasanna jokes that his career is a string of failed products. He cites Google Wave, Google+, a social startup called Secret and an email startup he built with Canva's co-founder. He says each taught him to avoid a class of error, and Cash App was the big success.

He closes by urging people in an era of change to focus on what matters to them. For him that is open source, open protocols and wider access, and technology should serve a purpose rather than be chased for its own sake. He asks listeners to demand more of employers, defaulting to open source and building for everyone, as AI companies lock themselves into walled gardens.