When Jack Dorsey cut more than 4,000 jobs at Block in February, the idea was that AI meant the company could operate with fewer than 6,000 people, down from more than 10,000, and still produce more.
Six months later, we have some numbers: Unsurprisingly, Block's adjusted operating margin reached 27%, up from 22%. Revenue and adjusted earnings beat expectations, and the company raised its full-year gross-profit forecast.
Square also shipped 130 features in the first half of 2026, compared with 40 in the same period last year. There were substantial releases: a new Register, drive-through support for restaurants, AI-assisted inventory management and Managerbot, which monitors sales, labor, stock and schedules. There were also smaller changes: adjustable receipt fonts, a sound for declined payments, negative-priced modifiers and an option to print a second receipt.
Much of business software consists of fixing minor irritations for particular kinds of customers. (I was recently in an Uber in SF and spent the entire ride hearing about how it's impossible to turn down audio instructions and safety guidance. Sometimes an issue for Trust and Safety becomes a new irritation for someone else.) However, counting each change as one "feature" makes unlike things look alike. The 130 new features also cover a period from January through June, so some predate the February cuts. Block has not disclosed Square's before-and-after staffing, whether its counting method changed or how many customers adopted the releases.
Block has offered better evidence elsewhere. In April, it said production code changes per engineer had risen more than 2.5 times since January. Builderbot and other AI tools were reviewing more than 90% of production change requests, and Builderbot was making 15% of production code changes with people intervening mainly for the final decision to deploy. Incidents following production changes were down more than 70% from a year earlier.
On the topic of Builderbot: It checks proposed changes against Block's architectural rules, security requirements and operating constraints. Different agents review different parts of a change in parallel. The system also examines incidents, internal announcements and messages, then proposes new checks for human approval. So Builderbot has a view across the codebase and changes its behavior as Block's people add knowledge and rules. As engineers use the tool, they are teaching it what Block should notice and protect.
That is still narrower than what Dorsey described in From Hierarchy to Intelligence. In March, he and Sequoia's Roelof Botha sketched a "world model" containing a continuously updated account of Block's work and customers. Eventually, they argued, an AI system could use that model to coordinate teams, identify what customers need and assemble Block's financial capabilities into new products.
Buzz, announced in July, is designed to collect more of the raw material. The open-source workspace combines chat, software repositories, workflows and AI agents. People and agents have identities and audit trails; messages, patches, approvals and automated actions go into the same searchable record. An agent can retrieve the history of an incident, review a patch or draft a release from the work recorded in project channels. Buzz could help Block build the company model Dorsey describes. For now, it is an early product,[1] not proof that such a model is deciding what Square builds. And to me, Square's feature log still looks like a conventional roadmap moving unusually quickly.
So has Block moved from Hierarchy to Intelligence? Not quite yet. It has built a serious system for producing and reviewing software, and it may be building the record from which a broader intelligence could learn.
2,200 PRs since March? This is a team moving very quickly, and the product is (perhaps un-) surprisingly feature-rich. ↩︎
