When AI agents join the group chat, unread counts stop working

Block describes Buzz as a shared workspace for humans and AI agents, not a chat app with one assistant bolted into the corner. An agent can have its own cryptographic identity and defined permissions. It can post, review code, run approved automations and join a conversation alongside people.

The app also combines surfaces that teams usually split across Slack and GitHub: channels, threads, direct messages, voice, media, repositories and automated work. Teams can use agents powered by different models, run their own instance or use Block’s hosted version. The code is available under the Apache 2.0 license.

TechCrunch’s early assessment is appropriately cautious. Buzz is available as a free desktop app for macOS, Windows and Linux, but the product is still early and its Git integration is not finished. A new startup with no settled chat stack may want to experiment. An established team should not move years of work because the launch video looks tidy.

Still, the product points at a real shift. AI agents are moving out of private one-person chat windows and into shared rooms where their output competes with coworkers, customers and deadlines for attention. That changes the interface problem.

In ordinary team chat, a message usually carries some social weight. A person chose to interrupt the room. Even a weak update took time to write. The unread badge is crude, but it roughly measures how much human communication arrived.

An AI agent has no such friction. It can post when a task starts, when a tool runs, when a source disagrees, when a test fails, when it retries and when it finishes. Each update may be accurate. Together they can bury the one sentence a teammate needed to see.

The obvious fix is not a smarter summary at the top of another feed. Summaries are useful, but they still make the person trust a compressed version of the noise. The stronger interface separates conversation from consequence.

Keep the channel for questions, decisions and disagreement. Put actual changes in a second, quieter timeline: document edited, customer record updated, automation completed, test failed, approval needed, nothing changed. A person returning to work should be able to scan that list in under a minute, then open only the threads that explain a surprising result.

First: did the agent only talk, or did it change something? Those states should not share the same visual weight. “I found a possible fix” is conversation. “I changed the billing rule” is an event.

Second: where did the change land? Name the file, record, board, repository or customer account. A vague “task complete” card forces the reader back into the transcript to discover what completion means.

Third: what is still waiting on a person? Do not mix a genuine decision with background progress. A single review item should not disappear beneath twenty cheerful status updates.

Fourth: what failed quietly? If an agent tried an automation three times and gave up, the room needs the failed end state, not a green-looking stream of activity. “No change made” is often the most calming sentence on the screen.

This is where AI assistants can return time instead of moving the work into review. The person should read fewer messages because the interface has already sorted speech from state change—not because the agent wrote a longer recap.

Ivy Chen would test Buzz in one channel with one repeated coordination problem: release notes, support escalation or a morning status check. Give one agent a narrow job, name the teammate who handles exceptions and decide which old check should disappear if the trial works. If everyone keeps the old process and adds an agent channel, adoption has created another surface to babysit.

Cass Bell is less impressed by visible activity. A channel can sound extremely productive while nothing outside the channel changes. She would count finished after-states: the issue closed, the document updated, the failed run marked, the human decision surfaced. Five agent updates with no changed system are not five pieces of progress. They are narration.

Those views pull in the same useful direction. The trial should remove a repeated check, and the interface should prove that it did. Message volume is not the evidence. A calmer return to the desk is.

Start with a channel people already avoid because catching up is tedious. Pick one outcome the agent can leave in a clear end state. Do not begin with “help the team.” Begin with “check these five sources every morning and mark changed, unchanged or blocked.”

Then cap public progress updates. The agent can keep a detailed internal record without posting every tool call into the room. In the shared channel, reserve interruption for a changed result, a failed result or a question only a person can answer.

Keep human and agent identities unmistakable, but do not stop at the avatar. Show whether an agent is speaking, proposing or reporting a completed action. A friendly name tells you who posted. It does not tell you what happened.

Finally, run the return-from-lunch test. Ask someone who was away for two hours to find what changed, what failed and what needs them. If they have to read the whole channel, the team does not have an AI collaboration system yet. It has a faster group chat.

Buzz is early, and that makes it a useful signal rather than an obvious migration. The interesting bet is not that agents deserve their own chat accounts. It is that workplace software now has to help people look away, come back and recover the state of work without giving the afternoon to an unread badge.