Block Builderlab Puts AI Agents in Shared Team Workspaces
- Cargo
- Machine Learning
- Shipped
- Shipper
- Sloane Kali Faye
- Readers
- 892
Block’s Builderlab connects named, auditable agents to team workflows, with cloud execution, model choice and security controls for delegated work.

Block’s Builderlab points to a shift in workplace AI: from tools that answer one person’s prompts to agents that take part in shared team workflows. Built around Block’s Buzz workspace, the system gives agents identities, lets colleagues tag them and records their actions. Its practical test will be whether teams can delegate meaningful work while retaining control over data and permissions.
Block first built these tools for its own operations as it pursued a model of working as an “intelligence company.” It is now extending them to other businesses through Builderlab, a set of tools for people and agents to work together. The approach builds on Buzz, Block’s open-source workspace for humans and AI agents, and the company has described Builderlab as a way to connect ideas with execution.
Agents become visible team participants
Many AI products are designed around an individual user. That can make the person a handoff point when a task needs to move from a private interaction into a team’s work, said Lucas Isaza, product lead for Buzz at Block. Buzz makes agents visible in a shared workspace: each has a name and avatar, can be tagged like a colleague, and leaves a record of its actions.
That design changes how people encounter the software. Rather than opening separate applications and moving results between them, a team can engage an agent in the same workspace where it discusses and tracks work. The identity and activity record also make an agent’s role easier for colleagues to inspect. These features don’t establish that an agent’s output is correct; they make its participation and actions more observable.
Block added plug-ins after Buzz’s first version launched in July. The plug-in architecture lets users reshape the interface, while the companion apps platform has drawn heavy use from nontechnical staff. Isaza cited account managers building personal dashboards as one example. That suggests customization is intended to reach beyond developers, though the source provides no usage figures or detail on how widely those tools have spread.
Open source, hosted services and model choice
Builderlab uses an open-source plug-in architecture that developers can customize. Companies can self-host the code, start with a free tier, or pay for remote agents, models and additional storage, Isaza said. Block’s proposed business pattern is familiar: make the software available openly and charge for hosted services that reduce the work of running it.
That structure gives organizations choices about deployment, but it also leaves practical questions for evaluation. Self-hosting may give a company more control over its environment, while hosted execution can shift some infrastructure work to a provider. The available details don’t specify service pricing, capacity limits or the operational effort required to manage each setup, so teams will need to assess those factors before committing.
Builderlab sits above the underlying models, according to Brad Axen, Block’s head of AI capabilities. Buzz can therefore work with agents from different AI labs, and Block can switch to a newer frontier model as one becomes available. This separation may make the workspace less dependent on a single model provider. It does not, by itself, show how well different models perform on the same task or how much work a model change might require from a team.
Delegation moves to cloud environments
Remote agents run in centralized cloud environments, which Axen said are easier to secure than laptops. The setup supports a different work rhythm: people can assign substantial tasks overnight and review the results later. That puts more weight on clear task boundaries, review practices and action records, since a person may not be watching each step as it happens.
Axen said some teams have shipped software releases 50% to 75% faster. The figure is a reported result from some teams, not a general benchmark, and no comparison method or task scope was provided. It’s a reason to examine delegated workflows, but it doesn’t establish that every organization should expect the same speedup.
Block’s design also aims to limit what an agent can do through technical permissions rather than relying only on instructions telling it what to avoid. Axen described “secure building blocks” that never grant permission for certain actions. He expects agents to need substantially less human steering within the next year, arguing that they already operate inside the conversation. That forecast remains an expectation; the relevant measure for adopters will be how often agents complete assigned work correctly and how much review it still takes.
Security follows familiar principles
James Nettesheim, Block’s chief information security officer and head of enterprise technology, said securing autonomous agents can build on established practices: least privilege, isolated environments and controls on outbound data. Builderlab gives each agent an identity owned by a person, runs agents in hardened cloud virtual machines and records actions in a cryptographically signed ledger that Block says cannot be altered.
Block is also developing information flow controls that tag data from sensitive sources. The aim is for security rules to travel with that data, preventing it from leaking into places where it shouldn’t go. This matters because an agent can encounter information across multiple tasks and tools. The approach will depend on how accurately data is classified and how consistently those controls apply as information moves.
Block already uses AI to review code and sort security alerts, Nettesheim said. He sees agents as a way for defenders to keep pace with attackers operating at machine speed. For organizations considering this model, the more immediate question is operational: can teams assign useful work, trace what agents did, and restrict access to the data and actions each task requires? Builderlab makes those controls part of the workspace design; adoption will turn on whether they hold up in day-to-day use.