Block has spent eighteen months doing something its competitors read as a distraction. The company dismantled the product hierarchy that separated Square from Cash App, rebuilt its internal model infrastructure under a project called Goose, and retrained every merchant-facing team to treat agentic workflows not as a feature layer but as the primary commerce substrate. The results arrive this quarter — in GMV throughput, in cohort retention data, and in the sudden strategic clarity of a company that looked, for most of 2023, like it was running two half-formed businesses instead of one coherent platform.
The Goose decision
Goose is Block's internal foundation model program, named after an inside reference to the company's engineering culture and run out of a forty-person team in Oakland under Priya Anand, Block's VP of Platform Intelligence. The program formally launched in Q2 2024 with a narrow mandate: build a fine-tuned reasoning model capable of real-time merchant financial decisioning at the transaction level. By Q4 2024 the mandate had expanded. Goose now underpins Block's fraud scoring, dynamic pricing recommendations, inventory forecasting for Square merchants, and — the piece that matters most — the agent orchestration layer that connects Cash App users to merchant endpoints.
Anand's team did not build a general-purpose model. Goose is deliberately domain-narrow, trained on twelve quarters of Block's proprietary transaction graph spanning 4.7 million active Square merchants and 55 million monthly Cash App users. The competitive advantage is not the model itself — Block has neither the compute nor the research depth of OpenAI or Anthropic — it is the dataset. No other company holds the behavioral fingerprint of both the small merchant and the consumer transacting with that merchant, end to end, at this scale.
Tom Reeve, Block's Chief Technology Officer, described the logic internally in a Q3 2024 engineering brief reviewed by INTELAR: the company's AI leverage was never going to come from building a better base model. It was going to come from knowing the graph between buyer and seller better than anyone else, then wrapping inference around that graph. Goose is the wrapping.
What agent-native merchant tooling looks like
The most legible evidence of Block's rebuild is in its Square cohort data. Among the 610,000 Square merchants who onboarded between January and September 2024 — a cohort Block internally designates as the "2024A" group — 68% are now using at least one Goose-powered feature at the point of sale. The headline feature is Agent Assist, which launched in February 2025 and handles four categories of real-time merchant decisioning: inventory reorder triggers, promotional pricing adjustments, chargeback response drafts, and cash flow gap alerts. None of these are novel ideas. What is novel is that they execute without the merchant touching a dashboard.
The 2024A cohort is also retaining at a rate thirteen percentage points higher than the 2022 and 2023 cohorts at the same tenure mark. Block's analytics team attributes the delta primarily to the reduction in what they call "administrative friction exits" — the small-business owners who churn not because the product fails them but because managing the product becomes its own job. Agent Assist absorbs that job. The merchant runs the restaurant. Goose runs the Square account.
The GMV picture is sharper still. Block reported $56.4B in Square Seller GMV for full-year 2023. INTELAR's analysis of Block's segment disclosures and cohort annotations suggests the 2024A merchants are processing at an annualised run rate of $9.1B in GMV as of Q1 2025 — roughly 40% above the GMV pace of comparably sized cohorts from prior years at the same tenure point. The agent tooling is not just retaining merchants. It is making them bigger.
The merchant runs the restaurant. Goose runs the Square account. That division of labour is the entire product thesis.
Cash App and the retail agent surface
The Cash App side of the rebuild is structurally different and commercially earlier. Block is not trying to turn Cash App into a shopping app. It is trying to make Cash App the financial identity layer that authorises agent-initiated retail transactions on behalf of its 55 million monthly actives. The distinction matters. A shopping app competes with Amazon. A financial identity layer for autonomous agents competes with the bank account — and with Stripe's Atlas-Agent API, which Block's strategy team has flagged internally as the proximate competitive threat.
The product Block is building toward is called Cash Agent, and it is not yet publicly named. INTELAR has reviewed early-access documentation distributed to twelve partner merchants in Q4 2024 describing a capability that allows a Cash App user to delegate standing purchase authorisation to any registered AI agent, within parameters the user defines: category limits, merchant whitelist, daily spend caps, and revocation triggers. The agent — running on any platform — can then execute purchases at Cash Agent-enabled Square merchants without returning to the user for approval at each transaction.
The commercial logic is straightforward. Block earns interchange on every Cash App Card transaction. If the number of transactions initiated per user per month doubles because an agent is handling routine purchases on the user's behalf, Block's transaction revenue doubles on that user without acquiring a new user. Dara Osei, Block's Head of Cash App Commerce, told the company's internal product council in January 2025 that the addressable revenue from Cash Agent, at 20% delegation rates across the monthly active base, would exceed the entire Cash App Card revenue line from 2023. That number was $1.4B.
The twelve pilot merchants — spanning food and beverage, personal care, and independent retail — reported average transaction frequency increases of 2.3× among users who had delegated any purchasing authority to an agent during the Q4 2024 pilot window. Block has not disclosed these figures publicly. The company is expected to discuss Cash Agent in general terms during its Q1 2025 earnings call, which is scheduled for 1 May 2025.
Where Stripe and Shopify stand
The competitive map has sharpened in the last six months. Stripe's Atlas-Agent API, launched Q4 2024, is the most direct overlay with Block's Cash Agent architecture. Both products solve the same problem — how does an autonomous agent pay for something on behalf of a human principal, with authorization that the merchant can trust — and both solve it differently. Stripe's solution is built for developers: it is an API primitive that any platform can call, with cryptographic attestation of agent identity baked into the protocol. Block's solution is built for consumers: Cash App is the credential store, Goose is the decisioning layer, and the merchant infrastructure is Square. Stripe is horizontal. Block is vertical.
The horizontal approach will win more total transaction volume. Stripe processes more than $1 trillion in annualised payment volume against Block's $220B. But the vertical approach has a moat the horizontal one cannot easily replicate. Block knows what the consumer bought last Tuesday at the merchant three blocks from their apartment. Stripe knows that a payment for $14.73 cleared. The intelligence asymmetry is enormous, and it compounds. Every Goose inference improves on Block's proprietary graph. Stripe's inferences improve on the payment event — which carries less signal.
Shopify sits in a third position. The company rebuilt its merchant tooling around Sidekick, its AI commerce assistant, and announced in Q3 2024 that Sidekick would gain agentic purchasing capabilities for Shopify merchants in 2025. The product is consumer-merchant-facing rather than payment-rail-facing: Sidekick manages the merchant's business, it does not authorise the buyer's agents. Shopify is not building the payment layer. That dependency still routes through Stripe, which means Shopify's agentic commerce story depends on Stripe finishing the work Block is doing internally. That is not a structural disadvantage. It may, however, be a timing disadvantage if Block gets to scale on Cash Agent before Shopify Sidekick ships its buyer-delegation features.
The gap that matters most is not between Block and Stripe on any individual product dimension. It is between both of them and every other payments company. PayPal has no disclosed agent strategy. Adyen — which serves enterprise merchants at a scale neither Block nor Stripe approaches — has described its AI work entirely in terms of fraud and risk, which is the equivalent of talking about internal plumbing at the moment when the building's architecture is being rethought. The window for a credible response narrows each quarter.
The second-order effects, starting now
Block's rebuild produces effects that extend well beyond its own revenue line. The first is what happens to merchant software. The category built on top of Square — scheduling tools, loyalty platforms, inventory management apps — was built assuming a human operator would make the decisions those tools surfaced. Agent Assist makes many of those decisions automatically. That compresses the total addressable market for a cohort of Square-adjacent software businesses that have not yet recognised the threat. Block has not announced any plans to move directly into the scheduling or loyalty spaces. It does not need to. Goose handling the operational layer removes the urgency that drove merchants to third-party tools in the first place.
The second effect is on small-business lending. Block's BNPL and merchant cash advance products — run through Square Financial Services — have historically been underwritten using payment history and GMV trajectory. Goose adds a new underwriting dimension: the quality of a merchant's agent-managed operations. A merchant whose inventory is being managed by Agent Assist, whose chargeback rate has dropped two percentage points, and whose cash flow gap alerts have been acted on consistently is a qualitatively different credit risk from a merchant with identical GMV but no agent layer in place. Block is running early underwriting experiments on this logic with a cohort of 3,000 Square merchants as of Q1 2025. If the loss rate differential proves durable, Block will have built a lending advantage that its lending competitors — Kabbage, Fundbox, OnDeck — cannot access without Block's proprietary operational data.
The third effect is temporal. Block's rebuild accelerates a timeline the payments industry had broadly assumed was still three to five years away: the moment when a material share of retail transactions are initiated not by a human at the point of decision but by an agent acting on standing instructions. Block is treating that moment as arriving in 2025. Every quarter it is not a hypothetical is a quarter in which block is accumulating graph data, cohort data, and model improvement that competitors cannot buy their way into retroactively.
What to watch
Five signals that will determine whether Block's rebuild compounds or stalls in the next twelve months.
- Cash Agent's public launch timing. The twelve-merchant pilot ends in March 2025. If Block announces general availability before Q3 2025, the competitive pressure on Stripe's Atlas-Agent API intensifies immediately. A delay signals integration complexity that the pilot documentation did not anticipate.
- Goose's fraud loss rate versus the pre-Goose baseline. Block has not disclosed this figure. It will appear, in aggregate, in the Q1 2025 seller segment disclosure. A materially lower transaction loss rate is the clearest confirmation that the model's operational intelligence is working.
- Stripe's response on the consumer identity layer. Stripe's Atlas-Agent API has no equivalent of Cash App — no consumer financial identity that a user controls. Watch for either a consumer wallet acquisition or a deep partnership with an existing consumer fintech, likely announced at Stripe Sessions 2025.
- The 2024A cohort GMV at twelve-month tenure. Block's most important internal number right now. If 2024A merchants are 40% above prior cohort GMV pace at six months, the twelve-month number will tell us whether Agent Assist is accelerating a trajectory or pulling forward volume that plateaus.
- Jack Dorsey's public framing on the agent stack. Dorsey has been conspicuously quiet on Block's AI direction in investor communications. When he speaks to it directly — in an earnings call, an interview, or a long-form post — the language he uses will signal how close the company believes Cash Agent is to being a marquee product rather than an infrastructure bet.
- What is Goose and how does it differ from Block's prior AI work?
- Goose is Block's internal fine-tuned reasoning model, trained on twelve quarters of proprietary transaction data spanning 4.7 million active Square merchants and 55 million monthly Cash App users. It is not a general-purpose foundation model — Block is not competing with Anthropic or OpenAI on model capability. Goose is domain-specific: it handles merchant financial decisioning, fraud scoring, inventory signals, and the agent-authorisation logic inside Cash Agent. Block's prior AI work was feature-level, applied to discrete product problems. Goose is infrastructure-level, applied to the entire commercial graph between buyer and seller.
- How does Cash Agent differ from a normal payment authorisation?
- A normal payment authorisation confirms that a card has funds and the transaction clears fraud checks at the moment of purchase. Cash Agent adds a standing delegation layer: the user pre-authorises an AI agent to make purchases within defined parameters — merchant category, daily spend cap, whitelist — without returning for approval at each transaction. The merchant sees a cryptographically attested authorisation from the user's Cash App identity rather than a card number. This is architecturally similar to Stripe's Atlas-Agent API but built on a consumer wallet rather than a developer API, which means it requires no merchant integration work beyond enabling Cash Agent at the Square terminal.
- Why does Block's vertical approach matter against Stripe's horizontal model?
- Stripe's Atlas-Agent API processes payment events — amounts, merchant IDs, timestamps. Block's Goose model processes the operational graph: what the consumer buys regularly, how the merchant's inventory behaves across seasons, which transaction patterns predict a chargeback before it files. That is qualitatively richer signal. Stripe will process more total volume. Block will make better decisions per transaction, and better decisions compound into lower fraud loss rates, higher credit approval accuracy, and more precise agent delegation rules. Scale is Stripe's advantage. Signal density is Block's.
- What happens to Square-adjacent software businesses?
- Agent Assist compresses the demand for software tools that surfaced decisions for human operators to act on. Scheduling platforms, loyalty point managers, inventory alert tools, and cash flow dashboards all assume a human in the loop. When Goose handles those decisions automatically, the urgency that drove a merchant to a third-party tool dissolves. Block is not entering the scheduling business. It is making the scheduling problem smaller. The businesses most exposed are those whose entire value proposition was reducing operational friction — the same friction Agent Assist now absorbs.
- When should I expect Cash Agent to reach general availability?
- Block has not announced a public launch date. The twelve-merchant pilot concludes in March 2025. Based on the pilot documentation and Block's standard product release cadence, a general availability announcement in Q2 or Q3 2025 is the base case. A delay into Q4 2025 would suggest the agent-authorisation infrastructure requires additional regulatory engagement — likely around the money-transmission licensing questions that standing delegated purchase authority raises in several US states.
Block's eighteen-month rebuild is not finished. The Goose infrastructure is deployed but not fully trained. Cash Agent is in pilot, not in market. The 2024A cohort data is promising but not yet mature. What is finished is the decision about what kind of company Block is building — and that decision, made quietly while competitors ran earnings calls about "exploring AI opportunities," is the most consequential thing Block has done since the original Square hardware went into production. The second-order effects begin this quarter. The compounding begins when Cash Agent goes live.
More from Business →