Video: "Buzz Just Fixed AI Agents... But It Has A Serious Flaw" by Julian Goldie on YouTube.
What Buzz is, briefly
Buzz is an open-source AI agent workspace created by Jack Dorsey's Block. It runs on the Nostr protocol — a decentralised, relay-based messaging system — and is designed as a place where AI agents and people work side by side inside channels, in the same way a team uses Slack. The agents are not add-ons or bots; they are first-class participants in the workspace, capable of taking tasks, reporting back, and handing off to other agents.
If you have not come across it before, we covered the initial release in an earlier article: Buzz is Jack Dorsey's free AI Slack — here is what it does and how it fits with Hermes Agent. The short version: Buzz is the workspace, Hermes Agent is a capable worker you can drop inside it. The two tools complement each other rather than compete.
The four agent types and where the fix lands
Buzz organises its agents into four types, and understanding them is necessary context for what the update changed.
Fizz agents are fully autonomous. You give them a task and they run it without asking permission at any step. They are the most capable type in terms of what they can complete unattended, and also the most likely to go wrong when the task hits an unexpected state.
Honey agents are approval-gated. Before each action, the agent pauses and presents what it plans to do. A human approves or redirects. They are slower but give you full visibility and control over what is being done on your behalf.
Bumble agents run on a schedule — a timer or cron-style trigger fires them, they complete their task, and they stop. Good for regular reporting, monitoring, or content jobs that need to happen at a set time.
Custom agents are user-defined combinations: you specify the behaviour, stopping conditions, approval rules, and triggers yourself.
The recent update focused primarily on Fizz. Fully autonomous agents have a practical problem: without proper stopping conditions, they can get stuck in loops, retry a failing action indefinitely, and consume API credits doing nothing useful. The fix added clearer loop detection and automatic escalation — when a Fizz agent cannot make progress, it now surfaces the block rather than silently grinding away. That is a meaningful change for anyone who has left an autonomous agent running overnight and come back to find it had spent most of its runtime retrying the same failed step.
What Julian Goldie identifies as the serious flaw
The structural problem is in how Buzz is built. Buzz runs on Nostr, which means every message, task update, and agent output is broadcast to a set of Nostr relays — servers run by volunteers across the internet. There is no central server that Buzz controls. Messages are delivered if the relays receive and retain them; if a relay is down or decides not to store a particular message, the message is gone.
For a casual workspace used by individuals, this is a manageable trade-off: the redundancy across multiple relays means most messages get through most of the time. But for a business workflow where you need a reliable audit trail — knowing that a task instruction was received, that the agent's output was saved, that a webhook fired — Nostr's architecture gives you no guarantees. There is no transaction log owned by your organisation, no replay mechanism, and no support contract behind the relays your workspace happens to be using.
This is not a bug the Buzz team can fix without fundamentally changing the architecture. It is a consequence of the design choice to build on a decentralised protocol. The benefits of that choice — no central platform, open data, no subscription fee to a vendor — come with this trade-off attached.
What this means for how you use Buzz
Buzz remains a strong choice for workflows where occasional message loss is tolerable and the open-source, no-cost model is valuable. Personal use, experimentation, low-stakes automation, and workflows that produce results you can verify elsewhere all sit comfortably within that boundary.
For workflows where reliability and auditability matter — scheduled financial reports, customer-facing automations, anything with a compliance requirement — the Nostr relay dependency is a problem you need a plan for before you deploy. Running your own private Nostr relay addresses some of the risk, but it shifts responsibility onto you and removes the decentralisation benefit that made Buzz free to use.
Julian's framing is consistent with how we think about this: Buzz is excellent at what it is, and the structural limitation is not a reason to avoid it but a reason to understand what it is and is not suited for before committing.
Where this connects to NordSys
Helping businesses choose the right agent infrastructure — and understand the constraints of each option before they matter — is part of what our AI Agents service covers. Whether that is a Buzz workspace for lower-stakes team automation, a Hermes Agent setup with full audit trails, or a hybrid stack that uses both tools for different jobs, the starting point is working out what your workflow actually needs from the reliability and auditability side. If you are at that stage, our AI Agents service is the right place to start the conversation.
See our AI Agents →