Goose is writing its own pull requests: how Block is trying to save a quarter of the week

Sonya Huang and Roelof Botha welcome Block CTO Dhanji Prasanna onto Sequoia’s Training Data. Earlier this year the company open-sourced goose, an extensible agent that works on your computer. It shows up in Jack Dorsey’s first cut of Bitchat, in side projects from people who are not engineers, and — the recursive bit — in the goose codebase itself. Roelof’s opening question is sharp. Block sells financial services and software to consumers and small businesses. Plenty of people say AI will make a company like that unnecessary. Friend or foe?

Dhanji gives two answers. The first is nuclear energy: nuclear medicine saves lives, bombs do not; it depends who holds the thing and what they do with it. AlphaFold is already on the good side of that ledger. The second answer is specific to Block. He has never thought of the company as a financial-services firm. It is a technology company that has always been early to whatever is new, then figured out how to serve customers with it. AI is a threat only if they are asleep at the wheel. The lineage he cites starts with the first Square reader, powered by the velocity of the card through the slot, and runs through their work on blockchain.

Roelof has been on the board since the end of 2010. Dhanji’s paper trail is almost as long: his first commit to the repo is 2011, with stints before a full-time return — January 2015 is only the most recent chapter. The two main pillars today are Square for merchants and Cash App for consumers. On top sit Tidal, the hardware wallet Bitkey, and a new mining box they call Rig, an attempt to build a powerful, energy-efficient home miner because Jack believes Bitcoin should be everyday money. Sonya notes that many AI-infra vendors began as crypto miners. Dhanji says the dual-use thought has not gone unnoticed, and that they are staying on mission for now.

Who drives AI at Block — Jack, Dhanji, or everyone? Before he took the CTO job, Dhanji sent Jack a long email: invest in AI, do it centrally, transform the whole company. Jack agreed 100 percent, flew to Sydney, and spent two days going through it. What was actually in the manifesto? “Hire someone, not me, who can be CTO and get our company transformed to using AI, because we’re well behind the eight-ball.” Jack followed half of that advice. The other half was appointing Dhanji anyway.

They moved in stages. First, eight special-project bets, two to five engineers each, a mix of Dhanji’s ideas and Hack Week leftovers; goose was one of them. Then they unwound the GM structure that, in Dhanji’s view, had locked Square, Cash App, and Tidal value inside silos, and folded the platform teams in with them. A functional org let them drive engineering excellence and unify policy. Roelof notes that org design is for a season. Dhanji agrees, and with standing: he helped spin Cash App out. He ran its engineering from about ten people to more than 200, working with Brian, then Cash App’s CEO, and Jack, so each business could keep a crystal-clear mission. In an era of weekly seismic shifts, he and Jack wanted depth and a single organizational focus instead.

Block has used machine learning for years. Dhanji draws a hard line. Traditional ML at the company was classification and clustering on the risk side — fraud, spam, abuse — and they got good at it. Generative AI, deep learning that can do more than those tasks, opens every vertical and every function. There is no panacea for putting it in the workforce. They tried “here is every AI tool, go tell us what works”; goose is now used by a majority of the company, and they still happily buy licenses for anything that beats it. The more structural move was to stop seeing Block as an enterprise plus some tools plus some products, and to see everything as a capability: take a payment, move bitcoin, buy a stock, file an issue, open a PR. An agent middleware layer — goose — sits on top, and UIs are evolving to talk to capabilities through it. He says that has already unlocked an enormous amount of value, and they are only at the beginning.

Goose, Dhanji says simply, is a general-purpose AI agent you download to a laptop, with a command line and a UI on the same guts. It is built on the Model Context Protocol, and Block was one of MCP’s earliest adopters — on the initial announcement, with a couple of extensions since. MCP, in his gloss, is a set of formalized wrappers that expose existing tools to an agent. Goose can already reach Gmail, Google Docs, Square payments, Snowflake, Looker, Tableau. Give it “a marketing report of how we did in Q3” and it pulls data, builds charts, ships a PDF or a Google Doc, and can email it, unsupervised. Sonya’s analogy — CLI like Claude Code, UI like ChatGPT — he accepts, with the caveat that both surfaces share the same MCPs. Engineers prefer the terminal; non-technical people prefer the UI, and they use it to build software too.

When he became CTO he rummaged in the back closet. Goose was one of seven or eight ideas. Brad Axen had been developing the thesis that agents would be how you actually get utility from AI. He was right. They ring-fenced him with six or seven people who have punched above their weight. The name is a *Top Gun* joke: Brad is a ringer for Goose. Roelof’s big-company fear is infanticide, firms killing their own ideas. Cash App and goose survived because Jack’s culture of experimentation gives people room. Plenty of bets do not work; Dhanji would add Bitkey and Proto to the short list of ones that did. At its core, goose pushes autonomy: the loop runs as far as it can, backs up when it hits a wall, tries another path. He says a lot of later agents learned that from them. It is also why people who cannot code can still ship something competent.

Early users worried most about safety: embarrassing email, deleted files, data they should not see. Dhanji’s experience is that the worry outruns the reality. Models are fairly cautious with tools, and goose adds a ladder: human-in-the-loop for every destructive action, then fully autonomous once you are comfortable. Even in autonomous mode it tends to flag a destructive step, and you can interrupt it onto another path. Most important, goose acts as you. It is not a wild robot in the data center. Access controls are the user’s, blast radius included. Sales does not see finance. A sibling, Headless goose, runs in CI: every vulnerability ticket from infosec, it tries a fix, and a human still has to read it before production. Then there is the usage Dhanji says you need a stomach for. One goose-team engineer lets it watch everything — Slack, Google Meet, the lot. He talks through a feature with a colleague; hours later goose has tried to build it and opened a PR, unasked. It breaks him out of flow if he is late and needs travel time. He says “let’s reschedule”; the calendar moves. “You have to have the stomach for it.”

The surprise, Dhanji says, has been non-technical people. Someone drops a Figma file and asks goose to make a working site. Someone holidaying in Paris has it build a map of sites and a traveling-salesman walk, then uses the app as a guide. Finance people have built treasury dashboards and one-click reporting tools. Jack’s first version of Bitchat — a decentralized Bluetooth chat and social app — was written with goose. They also use goose to build goose. The vast majority of its code is goose-written; they have almost fully bootstrapped it. Some human code remains at a complexity goose has not reached. The goal is a fully autonomous release that rewrites itself 100 percent from scratch.

Under the hood, goose is a pluggable provider system: any LLM that can tool-call. Sonya asks if she should think of it as an LLM-plus-tool-use loop. Dhanji prefers the image: the model is a brain in a jar, capable of nothing but chatting; goose is the arms and legs that take that thinking into Gmail and Snowflake. Tips for priming GitHub and Salesforce? Their approach has been not to over-engineer it. Let goose learn by doing. Recipes bake a workflow you like into a script you can share. They find goose more capable if you do *not* try to make a tool goose-friendly. It figures things out in surprising ways, and faster than a human would. Scaffolding you build for today’s model may be blown away by the next provider release. “You have to stop thinking like an engineer… and start thinking more like a data scientist.” He says that lesson has been hard.

Closed providers all have native tool calling and leapfrog one another. Open-source models generally emit text; some are fine-tuned toward tools. Block wrote Toolshim to adapt them onto MCP. People pick from a gateway of ten or twenty models, plus extra open-source plugins in the public goose — Dhanji built an embedded-model plugin on the plane over, “blindingly fast” on the latest MacBooks. Privacy-conscious users keep every token on-device with Qwen or DeepSeek. Coders lean Claude; GPT-5 is getting close. He would prefer every model open-source and open-weights, even though multi-trillion-parameter weights will not fit on a laptop. Block does not train LLMs. It does train smaller models for customer service and risk, and a speech-to-speech frontier model, research-only, that they plan to open-source with the findings. Models, in his view, should be a utility the way the internet was imagined to be.

Roelof puts the industry skepticism on the table: an MIT report suggesting very few Fortune 500 companies are truly benefiting from AI. Block has been at it long enough. What is the real impact? They track one weekly metric: manual hours saved by goose. It started at 0 percent. The year-end target is about 25 percent, and they are on it. The number mixes qualitative and quantitative signals. Engineers report eight to ten hours a week back from goose; Headless goose and the rest of the suite should push it higher. Dhanji grants the shape of the MIT complaint. LLMs are generalists. Ask about Soviet history and you get an instant answer; ask his wife, a neuroscience researcher, a narrow question about a type of dementia, and the model struggles. People with depth plus a model beat the model alone. The work is to find the model’s strengths and aim them. For Block that has meant automating workflows — getting “work about the work,” the drudgery, out of the way. The agentic middleware layer is the next bet, still early. Utility, he says, is still ahead.

Open-sourcing goose barely needed a debate. It is how Dhanji was hired: he used to work with Bob Lee, Block’s first CTO, and came in through open-source contribution. Roelof says a conversation with Jack years ago is why Sequoia started an open-source fellowship. Dhanji’s pitch to engineers who do not want to give code away is arithmetic. You can have 3,000-plus Block engineers on a problem, or 30,000-plus outside looking at it, contributing ideas, and benefiting too. Android is the proof: Block technologies run on close to four billion mobile devices. Goose should follow that tradition. Quality in open source is high because the bar to keep something maintained is high, all the way back to GNU.

Sonya asks whether open models will always sit a year behind the best closed ones. Dhanji sidesteps the frame. The way we use models is going to change. Today you have one or two models and a copilot on a laptop; goose’s core tool-calling loop averages two to three minutes a turn. What if fifty or a hundred instances worked for hours with each other — could they build something the size of Cash App? Then the contest is not open model X versus closed model Y. It is whether X is small and cheap enough to run on 50, 500, a thousand copies whose accumulated capability beats any single LLM. “Can an infinite number of ants build a spaceship?” He does not know. A hierarchical swarm is one picture: a strong model plans and reintegrates, and nano-tasks go to simpler models. It is a hot internal research direction.

Asked for a wish list, Dhanji always wants to move faster. Local tools have a responsiveness that builds momentum; he wants that for whole teams and org-sized initiatives. Roelof then names a tension the audience may miss: Block is remote-first, the CTO lives in another country and time zone, and that seems to fight speed. Dhanji will not pretend remote is perfect. Water-cooler serendipity really does accelerate work. What gets missed is the people you can only have because you support remote: leading lights who would never join from Sweden or Sydney. Cash App had distributed DNA from the start. You trade some velocity; you hire and retain engineers for six, seven, eight years in markets that do not have Silicon Valley’s competition. The Australian engineering office is almost a decade old. They have never looked back.

Sonya has just been with Anthropic’s Sebastian, who vibe-codes fixes himself when something in the app is off. Is Block in that era, and will non-technical people ever touch production? Dhanji unpacks two questions. Goose, he thinks, basically pioneered vibe coding, or was very early. He writes code every day, almost all through goose or rival agents he is evaluating; he rarely types by hand, maybe an edit or a commented-out line. A lot of engineers are still transitioning. Vibe coding is more effective on small, personal tools — dashboards, reports — than on ten-million-line legacy bases, a limit of context windows. Roelof presses: payments, crypto, security — surely an experienced developer still writes more compact, performant code than a model? Probably, in some very narrow cases. Even there, Dhanji wants the human to start from an LLM skeleton, like a sculpture’s armature. Models fail on proprietary APIs and complex internal frameworks. Humans are still called for on high-level architecture, race conditions, and orchestration across a topology.

Roelof, whose machine is locked down as a registered investment adviser, wants goose more accessible without admin rights. Dhanji offers browser and hosted experiments, and notes Sequoia’s own tech team already uses it. Safe mode exists. On the measurement Sonya wants: the most engaged engineers generate about 30 to 40 percent of the code they write in those gnarly legacy repos. On AI-first teams it is pretty much all vibe-coded. Goose itself: every PR that opens is written by goose. Other serious production apps are already deployed that way. It is still a while before every app is majority-AI.

Customer-facing generative AI is the other half. Dhanji thinks the productivity gains there may be even larger; the challenge is unlocking them with existing systems, which is why Block has pushed MCP with Anthropic and others. Square AI recently entered public beta: a goose-like bot that understands a merchant’s financials. You can ask for a Q3 bar chart, or something much more specific. A real wine-bar owner asked what they would lose by closing an hour early on Thursdays. The waiters who wanted the early close made a lot of their tips in that last hour. After talking to Square AI, they did not do it. The same treatment is coming for the rest of the product suite. Internally and in the products, everything is a collection of capabilities; agents sit on top; the hard part is evolving the interface so customers will actually engage.

Jack told the board that a company is capabilities and interfaces, and that people over-weight capabilities. Block has always put design and engineering on a pedestal. The first Square reader looked beautiful, worked simply, and hid enormous complexity. Cash App shows a single balance and orchestrates all sorts of money behind it. Voice, Jack has said, will matter as an interface; Dhanji agrees. The capabilities-plus-agent-middleware insight is one he and Jack talked through at the end of last year. Customers do not care that you have capabilities if they cannot use them to make a sale or move money. That is where the company is focused.

Three years out, Dhanji will not over-specify. A recommitment to being a technology company, to open source, to access and autonomy, should keep producing goose-like things. In three years agents may have evolved, or may not even be called agents. He wants Block at the front of that. Next year’s topic, he thinks, is unlocking utility. Many companies rushed in, failed to cash the value, and concluded AI is hype. Roelof quotes the pattern Roy Amara named: we overestimate new technology in the short run and underestimate it in the long run. His guess is a 2026 trough of disillusionment and an upside surprise by 2030. Dhanji is mostly with him, with a small chance that models keep improving at the current rate — diffusion plus transformers — and things take off. Companies that stay on their core reason for being, in Block’s case economic empowerment, will do well. Companies chasing dollars or hype will be left behind. Sonya is not in any trough. Goose writing its own PRs still looks to her like a magical technology in the hands of people who know how to build.