Nagent vs Buzz: a protocol and a product

Nagent is Multiplayer AI for end to end growth: a team of AI coworkers and your own people, working together in one workspace across marketing, sales and customer experience. Three things make it different. You approve the AI coworkers' work until they earn the right to act on their own. They carry the work all the way to pipeline and customers, not just content. And where your plan includes it, a Nagent marketer joins your team and owns the number with you.

When a company the size of Block decides to build a workplace for humans and agents, it is worth asking why it chose the shape it did. Block could have built a polished app and sold seats. Instead, as Block announced on 21 July 2026, it released Buzz as open source under the Apache 2.0 licence, built on Nostr, the open protocol for signed messages.

The choice tells you what Block thinks the hard problem is. In Buzz, every human and every agent has its own cryptographic keypair. Every message is an event signed with that key. The history of a room is a chain of those events, so any attempt to alter the past breaks the chain and shows. Agents are not bots bolted onto a chat tool; they are members, with identities as real as the people's. Block's bet is that as agents start doing real work, the thing organisations will need most is proof: proof of who said what, who did what, and that nobody has rewritten the record since.

The developer community agreed with the bet quickly. By mid September the repository had passed 33,000 stars and 4,000 forks, with more than 2,000 pull requests and 1,600 issues open. Few open source projects of any kind have moved that fast.

Nagent starts from a different question. Not how to prove what happened, but how to decide what should happen: which AI coworker is trusted to act on a customer, a budget or a live website, who approves the rest, and who answers for the result. Buzz is a protocol for engineers to build on. Nagent is a growth team you can hire, built for the people who run growth, most of whom will never generate a keypair or read a configuration file.

Open the Buzz desktop app and the first surprise is how familiar it feels. There is an inbox, a list of channels, direct messages, a search box. The second surprise is the Agents page.

Each agent is a card with a name, an avatar and the model it runs on. Different agents in the same workspace can run on different models. At the top right are two buttons that say a great deal about the product's priorities: one sets defaults for every agent, and one stops every running agent at once. Below, agent teams let you group agents and add them to a channel together.

Editing an agent reveals Buzz's most important design choice. An agent is a set of plain text instructions plus a harness, the program that actually runs it. The harness menu offers Buzz's own agent, Claude Code, Block's open source agent Goose, and Codex. In the workspace we looked at, Goose was not installed and the Codex adapter was missing, which is exactly what you would expect from software you assemble on your own machine.

Once configured, Buzz agents do serious work. In one channel of the workspace we looked at, a user asked an agent to build a small CRM with the Wasp framework and deploy it to Railway. The agent did, then announced the new repository to the workspace as a signed Nostr event, and before committing explained two changes it had made: it replaced a stock template README with real documentation, and it added an example environment file while confirming that the real one, with a database address in it, had been kept out of the repository. This is Buzz at its best: agents doing engineering work in the open.

In the same channel, the user asked the agent for a workflow that runs at eight every weekday morning and summarises what happened the day before. The agent created it, and the digest now posts on schedule: what shipped, what was decided, what is blocked and what is still open.

Several of these capabilities are still marked as experiments. Buzz's settings list workflows, described as automations defined in YAML with approval gates, alongside a git repository browser, an activity feed, forum channels and agents that manage their own profiles. The page is candid about their status: the features are functional but still being refined. Elsewhere in settings, a member can share their machine's local model with everyone on the relay, and each person's keypair and Nostr handle are fixed to the device.

The picture that emerges is of an exceptionally capable substrate for engineers. Buzz gives you identity, signed history, rooms, agents from any harness, git in the chat and workflows in code. What it asks in return is that you assemble the rest.

Imagine the head of growth at a consumer brand opening Buzz on a Monday morning. She is sharp, numerate and busy. She has a brand team, an agency, a paid media budget and a quarterly target. What does she find?

She finds a desktop app to install and a keypair tied to her device. She finds an Agents page with a New agent button, and behind it a blank instructions box and a harness menu. She finds that workflows are written in YAML and are still an experiment. If she wants agents for search, paid media, social, content and outbound, she will write each of them herself, choose their models and harnesses, and decide on her own what each is allowed to do. If she wants a CRM, the most instructive example we saw was an agent building one from scratch with a web framework and deploying it to a cloud host. That is an impressive demonstration for an engineer and an unsettling one for a marketer.

None of this is a flaw in Buzz. It is what infrastructure looks like. Linux does not come with a finance department either.

Now give the same head of growth a Nagent workspace.

There is nothing to install and nothing to assemble. Her growth team is already there: AI coworkers for search, paid ads, social, content, outbound and customer experience, organised into teams with an AI chief of staff called MIRA who plans the week and hands out the work. Her own people sit in the same workspace. The CRM comes included. Setup is a short conversation with MIRA, and the team starts working before it is even finished.

The question that matters to her is not whether a message was signed. It is whether an AI coworker should be allowed to do something at all. In Nagent, every AI coworker starts like a new hire, with everything it does waiting for a person. As it proves itself, Nagent recommends a promotion and she signs it off. A promoted coworker needs her less. After three months of clean work, for example, the AI coworker for search can earn the right to make routine website fixes on its own, once a person approves it.

And trust moves the other way too. In our demonstration workspace, the content coworker once let an unsupported product claim slip onto hundreds of pages. It lost the right to publish on its own, and the record says exactly what it must do to earn it back. Until then, every page it writes waits for a person. The decision and its history are written in plain English.

This is the layer Buzz's own community has started asking for. In August a contributor opened an issue in the Buzz repository requesting trust tiers for agents. Nagent already works this way, and goes one step further: trust is not only about who may instruct an AI coworker, but about how much it may do on its own, and it changes as its record changes.

Where her plan includes it, a Nagent marketer sits in the same workspace too, owning the number with her.

Buzz answers an important question for engineers: can we prove what an agent did? A growth team needs an earlier answer: before the action, has this AI coworker earned the right to take it, and who owns the result afterwards?

If your problem is growth and you want a team rather than a toolkit, see the plans on the pricing page, or book a walkthrough with the team. To see list prices at three team sizes, read the cost table in the Buzz (Block) guide.

• Nagent vs Buzz (Block): the sourced comparison guide, with dated screenshots
• Buzz (Block) alternatives: the other platforms in its category
• Buzz (Block) dossier: pricing, governance and evidence in one place
• The Earned Autonomy Ladder: How much an AI agent may do alone, decided by its record
• Anatomy of an Agent: A Configuration, Not a Program: The ten declared parts every Nagent agent is built from
• How to evaluate an agentic AI platform: Six criteria, three grades of evidence and one bounded pilot
• Autonomy is not a setting: Nagent vs NoimosAI
• Nagent vs Jasper: a hundred agents is not a team
• Nagent vs Okara: the founder is not a queue

Facts about Buzz are as its own pages and the reports below stated them, checked on 25 September 2026. Products change; the linked pages are the current word.

• Block, Buzz repository and README, GitHub https://github.com/block/buzz
• Buzz issue 6287 on configurable trust tiers, opened 19 August 2026 https://github.com/block/buzz/issues/6287
• Buzz issue 5372 on multi agent routing and observability https://github.com/block/buzz/issues/5372

Pratap Behera, Co-founder and Chief Executive, Nagent AI.

Pratap Behera is the co-founder and chief executive of Nagent AI. Before Nagent he led growth at BigBasket, where he helped build BBdaily into a USD 200 million business.

Published 25 September 2026, last updated 4 October 2026. Part of the Nagent vs series.