Forget OpenClaw and Hermes: why Buzz agents feel like real AI coworkers

Most AI agents today are great at talking but terrible at actually doing things. They live in separate dashboards, force you to copy and paste content everywhere, and rarely plug into your real workflows. Buzz takes a very different approach: it treats agents like real team members with hands, identity, and memory—able to write, get your approval, and then act in the real world.

From chatty prototypes to working AI coworkers

Previous tools like OpenClaw and Hermes Agent showed what was possible with AI agents: orchestration, tools, and self-improving skills. But they still felt like prototypes—powerful demos that didn’t fully live inside your day-to-day work.

Buzz pushes this idea into production. Instead of being just another assistant window on your laptop, an agent in Buzz behaves like a colleague in a shared workspace. It can join channels, see context, remember past conversations, and trigger real actions like publishing to LinkedIn when you give a simple thumbs up.

If you’re interested in how other platforms are making the same leap from prototype to production, it’s worth comparing this to the ideas in enterprise agent platforms built on Gemini.

Setting up a private social media channel

The core idea starts with a dedicated channel inside Buzz for social media. This is where all the work happens—drafts, approvals, and publishing receipts—without ever leaving the app.

You create a new channel, set its visibility to private so it’s just for you (or your small team), and enable the experimental Workflows feature in settings. Workflows are what let Buzz react to events in the channel—like new messages or emoji reactions—and then call external tools such as Zapier and LinkedIn.

Creating specialized AI agents: Wizard and Merlin

Instead of one giant do-everything agent, Buzz encourages you to create focused agents with clear roles.

Wizard: the research and voice agent

The first agent, Wizard, is responsible for understanding your brand and voice. You give it custom instructions and let it run on a powerful model (for example, via a Claude Max subscription). Then you drop it into your social media channel and assign its first task: research you properly.

Wizard goes off to study your YouTube channel, website, LinkedIn, X, and more. It then writes a detailed voice and style guide into the channel’s shared canvas. That canvas becomes the single source of truth for your tone, topics, and preferences, so any other agent you add to the channel can instantly write in your voice.

Merlin: the social media writer

The second agent, Merlin, is tightly scoped to a single job: writing LinkedIn posts as you. Its instructions are strict:

• Write LinkedIn posts only
• Use British English (in this example)
• Return just the post text, no commentary
• One post per reply

Once Merlin is added to the same social media channel, it automatically picks up all existing context, including Wizard’s voice guide. From that point on, you can tag Merlin in the channel and ask for posts about any topic, and it will write in your established style.

Why Buzz channels matter: shared context and identity

Buzz channels work like shared rooms where humans and agents collaborate. Everything that happens in a channel—messages, drafts, approvals, receipts—is visible and signed with a unique identity key.

Each agent has its own author ID, and so do you. Every action is cryptographically signed, so it’s always clear who (or what) did what. This identity layer is what makes secure approvals and audit logs possible. It’s also what lets workflows filter specifically on actions taken by a certain agent or by you.

Building workflows with YAML and Zapier

Workflows are where Buzz stops being a chat app and starts being an automation hub. They’re defined in YAML and scoped to a specific channel, and they react to events like “message posted” or “reaction added.”

To keep things simple, you can even ask an AI model to generate the YAML for you, then paste it into the Buzz workflow editor. From there, you wire Buzz to external services via webhooks—Zapier in this case.

Workflow 1: capturing every AI draft for approval

The first workflow handles draft capture. Its job is to grab every post that Merlin writes and store it in a Zapier Table for review.

Here’s how it works:

1. Trigger: a message is posted in the social media channel.
2. Step 1: Buzz calls a Zapier webhook, sending the message ID, text, and author ID.
3. In Zapier: the webhook triggers a Zap that filters messages so only Merlin’s posts continue (by matching Merlin’s author ID).
4. Storage: approved messages are written into a Zapier Table with columns like message ID and post text.

The result is a clean approval queue: every AI-generated draft is stored with a unique ID, ready to be published when you’re happy with it.

Workflow 2: publishing on thumbs up

The second workflow is where the magic happens. Instead of logging into a separate dashboard to approve content, you just react with a thumbs up emoji inside Buzz.

This workflow looks like:

1. Trigger: a thumbs up reaction is added to a message in the social media channel.
2. Condition: the workflow checks that the reaction was added by you (using your public author ID). Only you can approve posts.
3. Webhook: Buzz calls another Zapier webhook, sending the message ID and a flag that it’s been approved.
4. Receipt: Buzz posts a confirmation message in the channel saying the post has been published and approved by you.

On the Zapier side, the second Zap:

1. Receives the approved message ID from Buzz.
2. Looks up the full post text in the Zapier Table using that ID.
3. Uses the LinkedIn integration (logged in as you) to create a new share update with that text and chosen visibility (e.g., anyone or connections only).

End result: you see Merlin’s draft in Buzz, react with a thumbs up, and within seconds the post is live on LinkedIn. No extra apps, no manual copy and paste, and a signed audit trail of who approved what.

End-to-end example: from prompt to LinkedIn post

Once everything is wired up, the workflow feels surprisingly natural:

• You tag Merlin in the social media channel and ask for a post about giving AI agents real abilities in Buzz, not just chat.
• Merlin reads the channel history, understands the context, and drafts a short LinkedIn post in your voice.
• The first workflow sends that draft to Zapier, filters it by Merlin’s author ID, and stores it in a Zapier Table.
• You like the draft, so you react with a thumbs up emoji in Buzz.
• The second workflow fires, verifies that the reaction came from your key, and calls Zapier again.
• Zapier looks up the post by message ID and publishes it to LinkedIn using your account.
• Buzz posts a receipt in the channel confirming it’s live.

Because agents remember the conversation, you can then ask Merlin for a follow-up post referencing the first one, and run the same thumbs-up-to-publish flow again. Over time, the system starts to feel like a real social media team: one agent understands your brand, another writes, and you simply approve.

Signed events and built-in audit logs

A subtle but important feature of Buzz is that every event is signed with a key. Messages, reactions, and agent actions are all traceable to a specific identity. That means your approval isn’t just a click in some external dashboard; it’s a cryptographically signed event that lives right next to the content it authorized.

This matters for trust and accountability. You can always see which agent wrote which draft, who approved it, and when it was published. For teams, this becomes a built-in audit log of AI activity, not just a stream of anonymous automation.

Shared compute: community-powered local models

Buzz also leans into a different idea of AI infrastructure: shared compute. Instead of relying only on big cloud APIs, users can share their local models and hardware with each other.

When you enable the compute sharing feature, models running on your machine can be made available to other members of your Buzz community, and vice versa. In practice, that means:

• Your Mac Studio (or other machine) can host a local model that others can use as a backend for their agents.
• You can tap into models hosted by other community members without juggling API keys or worrying about token bills.
• As more people join and share, the community can collectively run larger or more diverse models.

This peer-to-peer style of AI compute is very different from the usual "pay per token" setup and fits nicely with the idea of agents as shared team members rather than isolated bots. It also pairs well with tools like MCP servers and orchestrators explored in setups such as multi-model Hermes agent stacks.

Why this feels like the “real deal” for AI agents

Putting it all together, Buzz shows what it looks like when agents stop being prototypes and start behaving like coworkers:

• They live in the same channels as you, with full context and memory.
• They have clear roles: one researches your brand, another writes posts, others could handle email, CRM, or support.
• Every action is signed and auditable, tied to a specific identity key.
• Approvals are natural—just an emoji reaction in the same place you’re already working.
• Workflows and Zapier bridge the gap between chat and the real world, pushing content straight to platforms like LinkedIn.

Instead of juggling dashboards and copying content between tools, you get a small AI team that drafts, waits for your signal, and then executes. OpenClaw and Hermes Agent helped prove that these systems could exist. Buzz demonstrates how they can live inside a real workspace, with identity, memory, and hands.