Buzz has the right idea. It just isn't ready yet.

I used the latest self-hosted version of Buzz for several days. Once it became clear that it was not ready for high-level work, I stopped.

I was rooting for it. My agents live in separate harnesses, use different models, and bring different tools and perspectives. I currently connect agents running through Claude Code, ChatGPT, Hermes, and OpenClaw across Telegram groups. Custom connectors, routing logic, and behavior rules hold the whole thing together.

Buzz could replace that mess with an actual product. The channel-specific context is a big part of the appeal. Each channel can have its own purpose, rules, and information instead of making every agent carry the same instructions into every conversation. My Telegram groups use broader shared rules, so that alone would simplify what I do now.

Same workspace, different agent access

Opening the same Buzz workspace on another computer gave me a different agent list. The agents still existed, and I could reach some through direct messages. In other channels and threads, I could no longer tag or call them.

It looked like a UI or state synchronization bug. I tried the available workarounds, but none restored full access reliably. Which computer I open should not change which agents I can use inside the same workspace.

Buzz 0.5.0 added a bring-your-own-harness interface for systems that support the Agent Client Protocol. (Buzz v0.5.0) That is useful, but the cross-device basics still have to work.

The usage made Buzz impractical

I connected Claude Code through my Claude Max 20x subscription. Buzz burned through my five-hour allowance in under an hour, with fewer than 30 messages and no coding or design work.

Across the Telegram groups I use, I routinely exchange hundreds, and sometimes thousands, of messages before reaching a limit. This was an informal comparison. Claude’s limits vary, and the workloads were different. The gap was still too large to dismiss.

Buzz’s public code points to one likely contributor: hidden prompt overhead. A request can include Buzz’s harness instructions, the agent’s system prompt and core memory, channel and workspace context, team instructions, and relevant thread or direct-message history. The visible chat is only one part of what the model receives.

Buzz’s base prompt says core memory, including identity, rules, and goals, is injected every turn. Add the other context and the result was enough bloat to slow Buzz down and make it impractical for heavy work. I noticed the same slowdown with GPT and Grok, although their limits were less punishing.

Buzz lowered its default agent parallelism from 24 to 10. Its tracker also includes reports of tokens being consumed when an agent’s answer never reaches the channel. I limited parallelism to one. The usage problem remained.

I kept having to call roll

I want agents to listen, contribute when useful, challenge bad recommendations, and stay quiet when they have nothing to add. Tagging every agent for every discussion turns me into a switchboard operator.

Telegram groups can explicitly allow bot-to-bot communication. I can ask a room to simplify an explanation without tagging anyone. Multiple agents respond, compare answers, and correct one another. When I address one agent directly, the others usually stay out unless they spot a serious mistake or risk.

The harnesses and shared rules create that behavior. Telegram gives the agents rooms where they can participate without individual summons. Buzz has workarounds, but none gave me the same flow reliably.

Why I use multiple agents

The Telegram groups I use are homemade versions of what Buzz is trying to become. Telegram supplies the rooms while each agent keeps its own harness, model, memory, and tools. Shared rules determine when an agent should speak, challenge another agent, verify work, or stay silent.

During a device sync problem, one agent proposed a change that could have wiped tens of gigabytes of device-only data. Two other agents examined the evidence, challenged the recommendation, and stopped the action.

Another agent once installed a web research capability and declared it finished. A second agent tested it, found that the tool worked while its safety review remained incomplete, and refused to approve it. The first agent corrected the work, and the verifier tested it again.

Those disagreements are why I use multiple models.

I would rather use Buzz than maintain homemade connectors forever. Buzz could provide the same collaboration in a cleaner place, and its channel-specific context could make it better than what I have now. It has to work reliably first.

What would bring me back

• Agents stay available across devices
• External harnesses become normal room participants
• Open conversations work without constant tagging
• Context and subscription usage become manageable
• Channel-specific context adds value without bloating every interaction

Most of the Buzz coverage I have seen on X and YouTube treats it as the next evolution in agent infrastructure. Maybe it will be. The version I used is not there yet, and this article is the warning missing from most of that hype.

My guess is that Buzz is a few months away from being reliable enough to use as a daily driver. Until then, there is no reason to migrate. Telegram groups, Slack workspaces, or Discord servers work better today. Keep using what works and check back later.