Lenny's PodcastDhanji Prasanna: How Goose agents save Block 10 hours a week
After Block reorganized from GM silos into one functional org; the open-source Goose platform now saves AI-forward teams 8 to 10 hours weekly.
CHAPTERS
- 0:00 – 5:37
AI productivity: separating hype from real gains at Block
Lenny and Dhanji open by contrasting skepticism about AI productivity with Block’s internal data. Dhanji shares reported time savings and frames AI value as a fast-moving wave that companies must continuously adapt to ride.
- •Block sees meaningful AI-driven productivity, not just pilots and prototypes
- •AI-forward engineering teams report ~8–10 hours saved per week
- •AI benefits are changing daily; success requires ongoing adaptation
- •Early examples hint at agents doing work proactively (not just chat)
- 5:37 – 7:34
The “AI manifesto” that got Jack Dorsey’s attention—and a CTO offer
Dhanji recounts how he noticed AI was missing from executive discussions and wrote an internal letter urging Block to become AI-native. That initiative led to deep conversations with Jack Dorsey and ultimately to Dhanji stepping into the CTO role.
- •AI wasn’t being discussed among top execs despite industry shift
- •Dhanji’s letter argued for central, company-level AI focus
- •Jack Dorsey spent extensive time discussing the thesis with Dhanji
- •The outcome: Dhanji offered (and accepted) the CTO job
- 7:34 – 8:53
Rebuilding a tech-first culture: sparks, programs, and hack weeks
Dhanji explains that the first phase of transformation was cultural: reasserting Block’s identity as a technology company. He describes concrete programs to reconnect top engineers, launch special projects, and revive company-wide hack weeks to restore experimentation energy.
- •Addressing ‘identity drift’ from tech company to fintech/financial services framing
- •Creating forums for top ICs to connect and align
- •Launching multiple small special projects (2–5 engineers each)
- •Reinstituting hack week to reignite experimentation
- 8:53 – 12:00
Why Block reorganized from a GM portfolio to a functional org
The conversation shifts to organizational design as a prerequisite for deep technical strategy and AI transformation. Dhanji explains how separate business-unit engineering orgs created silos and why central functional leadership enabled shared platforms, policies, and technical depth.
- •GM structure: business units operated like independent companies with separate engineering/design
- •Functional structure: engineers/designers unified under single leadership
- •Functional org enables shared AI/platform strategy and technical excellence
- •Reference to Apple’s functional reorg under Steve Jobs
- 12:00 – 15:27
Day-to-day engineering now: ‘vibe coding’ and background AI in CI
Dhanji contrasts AI-native teams building rapidly with minimal hand-written code against teams constrained by legacy systems. He also describes ‘always-on’ AI processes that run in pipelines to find vulnerabilities and propose patches while humans are offline.
- •AI-native teams using ‘vibe coding’ tools to build faster with less manual coding
- •Legacy codebases see slower gains but still benefit from AI assistance
- •24/7 AI processes in CI: vulnerability analysis, bug/ticket triage, patch suggestions
- •Adoption varies by proximity to tools and codebase complexity
- 15:27 – 21:31
How Block measures AI impact: manual hours saved and validation metrics
They dig into measurement: self-reported time savings plus corroborating metrics to avoid pure anecdote. Dhanji shares the company-wide estimate of 20–25% manual hours saved and explains why gains differ dramatically by team context.
- •Primary metric: ‘manual hours saved’ across the company
- •Validation via PR activity, feature throughput, and data science aggregation
- •Company-wide trend: ~20–25% manual hours saved (all functions)
- •Engineering gains vary: biggest in greenfield work, smaller in complex legacy systems
- 21:31 – 23:54
Goose: Block’s open-source AI agent built on MCP (giving LLMs ‘arms and legs’)
Dhanji introduces Goose as a general-purpose desktop AI agent that can orchestrate real tools, not just chat. He explains how Model Context Protocol (MCP) lets Goose safely wrap enterprise systems so models can take actions across software environments.
- •Goose is a downloadable agent with a chat-like UI and action capabilities
- •Built around MCP (Model Context Protocol) to connect tools and systems
- •MCP wraps systems like Salesforce/Snowflake/SQL to make them ‘orchestratable’
- •Goose is open source and extensible via writing new MCPs
- 23:54 – 28:41
What Goose can do in practice: reports, docs, integrations—and why open source matters
They walk through concrete workflows where Goose pulls data, writes code, generates charts, and ships outputs into docs or email. Dhanji explains why Block open-sourced Goose and how other companies (including peers) are adopting it.
- •End-to-end automation example: query data, analyze, chart, generate a report, publish/email
- •Pluggable model providers: Claude, OpenAI, or local open-source models
- •Goose enables rapid AI enablement of internal tools without waiting on vendors
- •Open-source philosophy: building tools meant to outlast and outgrow Block
- 28:41 – 32:16
‘Goose watches me’: proactive agents that anticipate work (screen + voice)
Lenny probes a striking anecdote: an engineer configured Goose to observe his screen and conversations, then proactively draft features and open PRs. They discuss the experimental nature, the productivity potential, and the risks/limits of autonomy today.
- •Goose can process screenshots to understand what’s on-screen
- •In extreme setups, Goose anticipates needs and opens PRs based on Slack/email discussions
- •Voice/listening capability enables hands-free scheduling and workflow automation
- •Current status: promising experiment; autonomy works ~partially and needs supervision
- 32:16 – 37:56
The near future of engineering: long-running autonomy, overnight experiments, and rewrite thinking
Dhanji predicts a shift from short ‘ping-pong’ chatbot sessions to hour-long autonomous work and overnight build cycles. He argues teams may run many parallel experiments and even consider rewriting systems more frequently as AI lowers the cost of iteration.
- •Moving beyond 5–7 minute sessions toward hours of autonomous agent work
- •Agents should work nights/weekends, building in anticipation of needs
- •Parallelizing experiments: build many options, discard most, keep the best
- •Provocation: envision deleting and rebuilding apps each release (‘RM-RF’ mindset)
- 37:56 – 44:06
Why human taste still matters: avoiding ‘AI slop’ and applying portfolio judgment
They discuss where AI still underperforms—especially in high-level judgment about what to build or whether to build at all. Dhanji emphasizes taste, process scrutiny, and leadership decision-making to keep automation aligned with real value.
- •Humans provide ‘taste’ and anchoring to prevent off-script outcomes
- •AI is weak at cross-portfolio judgment (what matters, what’s necessary)
- •Example tension: buy a vendor tool vs. build with Goose vs. change the process entirely
- •AI still lags in deep architecture/complex orchestration compared to senior humans
- 44:06 – 54:04
Hiring in the AI era: functional org impact, learning mindset, and AI-friendly interviews
Dhanji explains that AI hasn’t yet fundamentally reduced headcount needs for building products at Cash App scale. The larger hiring shift came from the org re-architecture, while interviews and evaluation now increasingly consider candidates’ comfort learning with AI tools.
- •AI hasn’t yet drastically changed staffing needs for large-scale products
- •Functional org reduced ‘engineers as commodity’ thinking and improved reuse/platform leverage
- •Hiring focus: learning mindset and eagerness to experiment with AI tools
- •Interviews evolving to allow/assess AI-assisted building, while still valuing fundamentals
- 54:04 – 59:54
Adoption lesson: leaders must use AI tools themselves (plus a receipts automation story)
Dhanji argues adoption is driven by executives modeling behavior—Jack and the leadership team use Goose daily. He and Lenny stress learning through solving real personal tasks, highlighted by Dhanji’s story of Goose collating therapy receipts via AppleScript and Notes.
- •Best adoption lever: leadership uses tools personally and publicly
- •Use-case-driven learning beats reading think pieces
- •Personal example: Goose gathered diverse receipts, converted/organized, synced, and shared
- •Goose can recover from failures by backing up and trying alternate approaches
- 59:54 – 1:13:38
Career and leadership lessons: Conway’s Law, ‘controlled chaos,’ and product over code quality
The conversation broadens to leadership and product-building philosophy. Dhanji shares counterintuitive views: code quality doesn’t determine product success, and controlled chaos can fuel creativity—provided reliability foundations are solid.
- •Conway’s Law: outcomes follow org communication structures; structure change is leverage
- •Counterintuitive: code quality and product success are often loosely coupled (YouTube vs Google Video)
- •Controlled chaos can energize teams; strict process can stifle creativity
- •Start small: Goose, Cash App, and Bitcoin experiments began as small prototypes/hack-week ideas
- 1:13:38 – 1:26:41
Failure Corner, lightning round, and closing principles: openness and purpose
Dhanji reflects on multiple failures (Wave, Google+, Secret) and frames them as sources of humility and learning. The episode closes with rapid-fire personal recommendations and a call to demand openness—especially as AI platforms trend toward walled gardens.
- •Failures as teachers: strings of misses preceded Cash App–level success
- •Humility, listening, and avoiding repeated failure modes
- •Lightning round: books, shows, Steam Deck, life motto about changing what drains you
- •Closing: push for open source/protocols and purpose-driven tech in the AI era
