Interview, Fireside Chat
Block CTO Dhanji Prasanna: Building the AI-First Enterprise with Goose, their Open Source Agent
- Block CTO Dhanji Prasanna defines AI at Block as a "friend" contingent on developer intent, drawing a parallel to nuclear energy where the outcome depends on application (e.g., medicine vs. weapons).
- Block views itself as a technology company rather than a financial services firm, positioning AI as a tool to strengthen its business rather than a disruptive threat, provided the company remains active in adoption.
- Block's organizational structure shifted from a General Management (GM) silo model to a centralized functional organization to facilitate deep technological focus and unify engineering policies.
- Generative AI is distinguished from Block's historical Machine Learning (ML) by its capacity to handle every vertical and function (beyond traditional risk/fraud classification) via deep learning.
- Block's AI transformation began with a "manifesto" email from Prasanna to CEO Jack Dorsey advocating for central investment, leading to the hiring of Prasanna and the formation of special project teams.
- Block adopted a "capabilities-first" architecture, treating all tools (payments, invoicing, issue tracking) as modular functions exposed via an agent middleware layer called "Goose."
- "Goose" is an open-source, extensible AI agent built on the Model Context Protocol (MCP) that allows users to orchestrate workflows across disparate systems (e.g., Snowflake, Looker, Gmail) via natural language.
- The Goose project originated from an internal hack week idea by engineer Brad Axon, who was "ring-fenced" with a team of 6–7 engineers to develop the concept.
- Goose utilizes a "recipe" feature where successful workflows are baked into shareable scripts, allowing the agent to learn from execution rather than requiring manual tool configuration.
- Block's internal safety model for Goose employs a laddered approach ranging from "in the loop" (requiring human approval for destructive actions) to fully autonomous modes, with access controls strictly limited to the individual user's authorization level.
- "Headless Goose" is a specialized extension running in CI/CD pipelines that automatically attempts to fix InfoSec vulnerability tickets, subject to mandatory human audit and review before production.
- Internal metrics indicate Goose is projected to save 25% of manual hours by year-end, with highly engaged engineers currently generating 30–40% of their code via AI.
- Jack Dorsey's BitChat decentralized application was initially built using Goose, highlighting the agent's capability to bootstrap complex software without traditional coding knowledge.
- An engineer demonstrated extreme autonomy by having Goose monitor his communications (Slack, Google Meet) and automatically open Pull Requests for feature ideas or reschedule meetings without explicit prompts.
- Goose supports a pluggable provider system, allowing the use of various LLMs (e.g., Llama, DeepSeek, Qwen, GPT-5) and utilizing a "tool shim" to enable tool calling for open-source models lacking native support.
- Prasanna predicts a future shift from single-agent "co-pilot" interactions to "swarm intelligence," where thousands of smaller, cheaper agents collaborate to solve complex tasks like building Cash App.
- Block remains committed to open source, viewing Goose as a contribution to uplift the industry, with a belief that open models will eventually match closed models through cumulative swarm capabilities.
- Square AI, a customer-facing variant of Goose, entered public beta to allow merchants to query financial data and simulate business scenarios (e.g., testing revenue impact of store closure times).
- Block maintains a remote-first strategy to access global talent (e.g., Australia, Sweden), accepting minor trade-offs in serendipity for the benefit of retaining specialized engineers in non-Silicon Valley markets.
- "Vibe coding" is now standard practice for CTO Prasanna and AI-native teams, though manual coding remains necessary for complex legacy systems and high-level architectural orchestration due to LLM context window limitations.
- Block forecasts 2026 as a potential "trough of disillusionment" where hype fades, followed by a resurgence of value extraction by 2030 as companies focus on utility rather than experimentation.
- Future product strategy emphasizes evolving user interfaces to expose underlying capabilities, treating AI agents as the primary middleware layer for all digital interactions.