How Block rebuilt itself as an AI-native company at scale

Executive overview

Block restructured its entire engineering organisation — from siloed business units into a single functional org — and made AI adoption a company-wide priority. The result: engineering teams using Goose report 8–10 hours saved per week, and the company estimates roughly 25% of manual hours saved across all functions.

The biggest productivity unlock wasn't an AI tool. It was the org structure change. Functional alignment let Block share platforms, move engineers across teams, and drive technical depth in a way the GM model never could.

Building like a technology company first — not a fintech, not a financial services firm — is the prerequisite for everything else.

Restructuring from GM to functional org

• Block operated as a portfolio of near-independent companies: Square, Cash App, Afterpay, Tidal each had their own engineering, design, and leadership.
• No shared technical strategy, no common tooling standards, no cross-team mobility.
• Dhanji's first move as CTO: reunify all engineers under one leader, all designers under one leader.
• This is what Jobs did returning to Apple — not following a playbook, but rediscovering the same truth.
• Conway's law is real: you ship your org structure. Changing outcomes requires changing the structure of relationships between people.
• The transformation was painful. Momentum existed in each silo and not everyone agreed.
• Result: shared language, shared policies, engineers who can move to areas of need, and a coherent technical strategy.

Goose: Block's open-source AI agent

• Goose is a general-purpose desktop AI agent built on the Model Context Protocol (MCP) — Anthropic's open standard for wrapping tools so LLMs can manipulate them.
• Before MCP, LLMs could only chat. Goose gives the model "arms and legs" to act in the digital world.
• Use cases range from organising photos to writing and running SQL, generating charts, producing reports, and emailing them — all autonomously from a single prompt.
• Goose is fully open source. Any company can download, extend, and write their own MCPs for it.
• Provider-agnostic: supports Claude, OpenAI, and open-source models via Ollama.
• Competitors, partners, and mid-tier tech companies (including Databricks) are actively using it.
• Named after the Top Gun character — the engineer who built it reportedly looks exactly like him.

Measuring productivity gains

• AI-forward engineering teams self-report 8–10 hours saved per week.
• Block validates this with PR counts, feature throughput, and a data-science formula combining multiple signals.
• Company-wide (including legal, risk, support): roughly 20–25% of manual hours saved.
• Gains are highest on greenfield code bases; complex legacy code bases see less benefit today.
• Background AI processes run 24/7: analysing vulnerabilities, reading bug tickets, opening patches while engineers sleep.
• Key framing: "This is the worst it will ever be. This is now the baseline."

Who benefits most from AI tools

• Senior engineers are relieved — AI handles tasks they've done a thousand times.
• Junior engineers adopt fast, like early smartphone users: blitzing through things without overthinking.
• The biggest surprise: non-technical staff. Risk, legal, and enterprise teams building their own internal tools — compressing weeks of work into hours.
• Example: the enterprise risk management team built a self-service risk system using Goose, without waiting for an internal apps team.
• This reduces the volume of one-off requests hitting engineering — but also increases appetite for software across the company.
• Analogy: wider highways attract more cars. More people building software means more software to build.

Gosling and automated mobile testing

• Gosling is Goose for mobile — operates Android at the OS level using the accessibility API.
• Replaces armies of QA contractors clicking through every screen manually.
• Bakes UI tests into automated runs and produces reports at the end.

The engineer with Goose watching his screen

• One engineer on the Goose team runs Goose in continuous screen-watch mode, always on.
• Goose listens to his Slack messages and calls, interprets discussions about potential features, then opens PRs on its own a few hours later.
• It also reschedules meetings, nudges him when he's running over, and takes other proactive actions he never explicitly programmed.
• Not a production feature yet — an experiment from someone on the core team — but illustrative of the direction.

The future of how engineers work

• Current "vibe coding" is too ping-pong: prompt, wait 3–5 minutes, nudge, repeat.
• Block is pushing Goose sessions from a median of 5–7 minutes toward hours-long autonomous runs.
• Target: describe multiple experiments in detail, go to sleep, wake up with all of them built.
• Dhanji's working practice: write code daily, but throw away enormous amounts of it. Delete whole systems that don't feel right and restart.
• Aspiration: every release deletes the entire app and rebuilds from scratch — not possible today, but directionally correct.
• LLMs sitting idle overnight and on weekends is wasteful. They should be building in anticipation of what's wanted.
• Self-improvement experiment: Dhanji gave Goose a wish list of 10 improvements to itself. Success rate ~60% on well-described features.

Hiring and what to look for

• AI hasn't yet changed how many engineers are needed to build a product at Cash App scale.
• The functional org changed hiring more than AI did — headcount is no longer treated as a commodity.
• Now optimising for depth, modularisation, reuse, and platform investment rather than bodies per feature.
• Hiring signal: a learning mindset. Eagerness to embrace tools, not mastery of them on day one.
• Jack Dorsey's framing: Block should be a "learning first" company — what can we learn from every experiment?
• Interview approach evolving: asking candidates to vibe code, not just whiteboard pseudocode.
• Critical thinking and technical depth still matter more than being AI-native on arrival.

Build vs buy in the AI era

• AI makes it tempting to build every internal tool in-house instead of buying SaaS.
• The trap: you still own the maintenance, the edge cases, the long tail of support.
• Rule: keep coming back to core purpose. Block's is economic empowerment — anything that doesn't serve that is a distraction.
• AI judgment gap: AI can propose building a tool or buying a vendor solution, but it can't step back and ask whether the underlying process is even necessary. That's still a human call.
• InfoSec example: teams tie themselves in knots securing something, when the right answer is often "don't build that at all."

Lessons from building Google Wave, Google+, and Cash App

• Code quality has almost nothing to do with product success. YouTube was acquired with a notoriously messy codebase — videos stored as blobs in MySQL, Python stack running slowly — and became Google's most successful product.
• Google Video was technically superior: more formats, higher resolution, longer videos. YouTube won anyway.
• Google Wave: 70–80 engineers building before it had real users. Tried to be everything to everyone. Failed.
• Google+: another epic failure. Secret (anonymous social app): burned bright, then blew up. An email startup with the Canva co-founder: fizzled.
• Cash App: started as a hack week idea, grew from 10 engineers to 200+, from near-zero to 10–20 million users. Controlled chaos, not strict process.
• "A lot of engineers want to refactor. All this code could be thrown away tomorrow. Focus on what you're building and whom you're building for."

Start small — always

• Goose started as one engineer's side experiment. Brad (the creator) believed in agents before the buzzword existed. Shared with Databricks and Anthropic, built momentum.
• Cash App started as a hack week idea.
• Block's first public-company Bitcoin product: a hackathon team of three — Dhanji, Jack Dorsey, and one other engineer — bought a coffee at Blue Bottle with Bitcoin over Cash Card.
• Counter-example: Google Wave started big, failed. "If you're making a cup of tea, just make the cup of tea."

Leadership lessons

• Conway's law is more powerful than most people appreciate — structure determines outcomes more than tools or intent.
• You only hear about it when things are going wrong. Make time to step back and look holistically; don't mistake silence for success.
• Question base assumptions constantly. "Should we even build this at all?" is the most important question, and the one AI can't answer.
• Don't accept what's meted out to you. If you're not waking up energised about your work, change something.
• A year from now, what feels like a monumental decision will look trivial. Use that as permission to act.

On open source and what to demand from AI companies

• Block believes in open protocols and open source as a core mission — not just a tactic.
• Goose is free because Block is built on open source (Linux, Java, MySQL) and feels the obligation to give back.
• The internet was built as a promise of open sharing. AI should fulfil that promise, not close it off in walled gardens.
• Demand openness from employers, teams, and AI companies. Technology is here to serve people, not capture platform value.