Will Jack’s Buzz Attack Sting Slack?

_{Guest contributor: Alex Salkever is the Editor-in-Chief of the Agentic AI Foundation and the Linux Foundation, where he is responsible for storytelling across agentic AI and all of open source.}

Block founder and CEO Jack Dorsey tossed a jar of stinging insects into the agentic conversation with Buzz, an open-source workspace where people, code, and AI agents share the same channels. Millions noticed, and many immediately took issue with the data policies around its hosted version. Fair enough. This is the agentic era: everybody arrives carrying a pitchfork, a threat model, and several strongly held opinions about data sovereignty and risk.

But Buzz is interesting on a few deeper vectors. It is less a Slack killer, at least for now, than a test case for a set of enterprise-AI ideas that once sounded fringe but are starting to appear more viable. Specifically, what does work look like when we not only bring ourselves but our agents and all their context along for the multi-player ride? What happens when we can pool compute to host larger models in a hybrid-distributed fashion that might make it much easier to run near-SOTA on fractional hardware? And, perhaps most importantly, what happens when we give agents something much closer to a true identity rather than an API key? Enterprise chat’s network effects are strong. It’s one thing to ask your team to try Buzz and entirely another to ask your customers or your suppliers. Then again…remember Yammer?

We have had BYOD, bring your own device, and BYOC, bring your own cloud. BYOCLI, bring your own terminal. Now comes BYOA: bring your own agent but with a buzzy twist. In the Buzz formulation, an agent is not a glorified slash command that appears when summoned, performs a trick, and disappears. It can live in the shared workspace, take part in a thread, work on code, participate in approved workflows, and hand work to a human or another agent. Slack announced something similar with Claude operating in threads. But Buzz offers something more radical by allowing each participant to bring and share their own agents. That sounds simple, but it is a meaningful change in posture with loads of challenges around audit trails, context exposure, and more. Slack was built for people talking to people, occasionally interrupted by a cheerful bot announcing that Jenkins had caught fire. Buzz is built on the premise that the agents are on the team and that we can introduce them to each other.

That may turn out to be either the future of work or a very efficient new way to create 47 unread channels where agents yammer on. I couldn’t help myself. Probably both. But once teams run multiple specialized agents, which can then spawn their own work interactions, humans become the sticky middleware bottleneck if the agents cannot share context, work, and handoffs in the same place. Underlying all this is a simple truth: BYO CLI gave individuals a choice of agent. BYOA lets those choices enter a shared work graph, where they can actually layer together.

The old saw about the two ways to make money in tech is playing out yet again, right before our eyes. We are multi-player creatures and expect our tools to be multi-player. In that realm, we bring our own stuff. How you bring a team of agents and then interact with other people and their teams of agents is a novel canvas with many challenges. Buzz tries to put that collaboration in one space. Channels, threads, code repositories, workflows, and agent activity sit together rather than bouncing between Slack, GitHub, a coding-agent terminal, and several browser tabs that have become emotionally significant. A feature branch can become the conversation around that feature branch. The context is not pasted from one system into another and quietly lost along the way. Of course, this is really a work in progress, but that vision has hewn into view.

The big question is not whether companies, and the humans who still run them for now, will want agentic collaboration. They plainly do. Multiplayer is inevitable. Rather, the real breaking point is whether legacy systems can sufficiently adapt to accommodate the new needs of workflows, workspaces, and processes that depend on agents having an elevated status somewhere between FTE and trusted automaton.

Buzz is model- and agent-agnostic. Teams can bring their own agents, use their own model providers, and connect harnesses such as Codex, Claude Code, Hermes, or Block’s own Goose. This is a clever way of reducing barriers to entry and not forcing an early choice. Think of Buzz as the flexible overlay for all the other stuff you are doing in agentic and you kind of get the idea. By allowing everyone to not only BYOA but BYOH—bring your own harness—Buzz allows everyone to party on some level without forcing painful tradeoffs.

Where this gets truly radical is the Buzz Mesh vision of model sharing. Here’s the backstory: the first generation of enterprise AI was built inside vendor silos. This agent belongs to this model provider, runs in this cloud, speaks through this chat surface, and has access to whatever the platform permits this quarter. We are well past that point, with most developers having access to multiple models through an IDE, routing layer, or some other option. But, as we are all hearing, the piper has come for his pay. CFOs are cracking down on untrammeled usage of big SOTA, and the large model providers themselves are trying to improve efficiencies.

However, as Hugging Face’s latest open-source report points out, smaller models are downloaded and deployed at much higher rates than very large systems. At the same time, we are seeing neolabs such as Poolside and Thinking Machines put out models that are small enough to run effectively on a powerful desktop or AI appliance like a DGX Spark, but come close to matching SOTA performance. Google’s Gemma 4 has blown past 200 million downloads and is perceived as a stellar local model that can even run on phones. The local penalty is shrinking for many use cases.

But Buzz’s proposed play is not merely “bring your own model.” Plenty of tools can already say that. Rather, the play is that, once you bring your own agent into the shared workspace, someone else in that workspace can lend it more muscle. Under the Mesh design, an agent’s model requests could run on another authorized member’s machine, with model traffic traveling directly between the peers rather than through the Buzz relay. In other words, the person with the GPU box, AI appliance, or slightly alarming home lab can share inference capacity without becoming the company’s unofficial cloud provider.

That turns BYOA into something more interesting. The old choice was bring your API, or bring your small model. Now you can bring your agent, your harness, and your preferred model strategy—and let the hive help power it. Add the rising tide of data and AI-sovereignty concerns, and a locally hosted AI-first agentic collaboration platform with BYOA, BYOM, and BYOH, and suddenly a bit of shared compute starts looking pretty dang spiffy.

This loops us back to where we started originally, which is Slack. Add in Discord and Microsoft Teams, and you round out the triumvirate. The reality remains the same: every chat platform has its own strong network effects that are hard to uproot. Specifically, your users, customers and partners are all in the chat and won’t just fly over to Buzz on a dime. (Integrations have never fully solved this issue). Buzz faces an upward battle, going toe to toe with incumbents. That said, it may have an interesting backdoor. Developers and AI aficionados are clearly promiscuous, hopping across agents, models, and harnesses with abandon. Every weekend, there's something new to try, so why not try everything out? For Buzz, this offers the opportunity to stake out a beachhead of users and, over time, grow to encompass more functionality so that the embrace becomes both irresistible and invisible at the same time. To be clear, Block has never positioned Buzz as something even remotely related to Slack, but spend any time with Buzz, and the overlaps are striking.

As we move to a multiplayer AI universe, all of the threads that Buzz pulls on will likely be tugged by other products and examined by the incumbents. The questions Buzz asks are the right ones. Namely, how do we set up systems where delegated agents from multiple parties can interact with autonomy and security while respecting privacy and information access restrictions? Slack, a product juggernaut as sticky as they come, is not going anywhere. Buzz may not hit critical mass. But the agents will want to talk to agents, and I’ll want my agents to talk to your agents, and so the buzz is coming, one way or another.