Every product that puts AI agents to work in a company picks one of three places for them to live: a chat box, a workspace the agents are members of, or the machine their owner already works on. The choice decides who can reach the agent, whose credentials it uses, and what is left behind when the work is done.
2026 has produced good examples of all three, so this is a map rather than a pitch. Where muti sits is stated plainly at the end; the rest is the landscape as it stands in September.
Three shapes: a single thread in a chat box, a workspace with agents as members, and a room around the machine the agent runs on.
Shape one: the chat box
The agent lives in a conversation. You open ChatGPT, Claude or Gemini, describe what you need, paste in what it needs to see, and copy the result out. This is where almost everyone started, and for a single person answering a single question it is still the fastest path.
What it cannot do is hold work. The thread is yours alone, so a colleague who needs the same thing starts over. Context arrives by pasting, which means the agent sees a snapshot rather than the live file. Nothing the agent did is addressable afterwards: there is no id to quote, no record of who approved what, no artifact anyone else can open. The chat box is an excellent interview and a poor workplace.
Shape two: a workspace where agents are members
The 2026 answer. The agent gets an identity, joins rooms or channels alongside people, keeps memory across conversations, and is addressed the way a colleague is. Four products show how differently that idea can be built.
Block's Buzz, released under Apache 2.0, describes itself as "a workspace where humans and agents build together, on a relay you own". It is a self-hosted Nostr relay: every message, review, workflow step and git event is a signed entry in one log, and agents hold their own keypairs rather than a permission flag. A bridge called buzz-acp lets Goose, Codex and Claude Code join as members.
Kylon takes the opposite route: a finished, hosted product. Its first layer is the Room, holding files, people, agents and small business apps. Agents are, in its words, "members, not features", each with a role and durable memory, and they only see the rooms they belong to. On top sit spreadsheet-like Apps with their own databases, more than three thousand integrations, Skills, Workflows, and a memory tree that compresses long history into something searchable.
Coze 3.0 puts several people and several agents in one project, with files and outputs staying in the project and work split by mentioning different agents. Doubao's work partner skips the new workspace entirely and gives the agent an organisational identity inside Feishu, where the company's messages, documents and permissions already are.
The trade is the same in all four: you gain a place where work accumulates, and the agent runs where the workspace runs, on credentials the workspace holds.
Shape three: the machine you already work on
The third answer keeps the agent exactly where it is. Claude Code or Codex stays on one person's laptop, with the repository already checked out, the tools installed, the browser logged in, the subscription already paid. What gets added is reachability: a room around that machine so other people can send work to it.
This is what muti does, and what Alook and Tutti do in their own ways. Nothing is copied anywhere. The daemon on the machine polls out over HTTPS, runs the request locally and streams a timeline back, so the room holds messages, tool-call summaries, approvals and artifacts, never files or keys. See what the daemon uploads.
The cost is availability: a sleeping laptop is a sleeping agent, which is why requests queue in an inbox and the daemon installs as a login service.
A fourth shape arrived in September: the personal agent
Between 11 August and 29 September, SpaceXAI's Grok Bot, Meta Muse, Manus Cue and OpenAI dots all shipped the same idea: an agent that belongs to one person, lives on a cloud computer of its own, remembers across conversations and acts on a schedule. Tencent's WorkBuddy and Alibaba's QwenWork and QoderWake had reached the same place from the other side, running on the user's own computer and taking work from WeCom and DingTalk groups. Personal agents answer a different question from the three shapes above, what can one agent do for me, and the cloud ones mostly leave "who else can use it" unanswered. The map is in personal agents in 2026, compared.
How the three compare
Which shape fits which problem
Pick the chat box when the task is a question and you are the only person who needs the answer.
Pick a workspace when the work is the company's rather than one person's: a shared CRM the agent maintains, recurring workflows, a record several departments read. Choose a self-hosted one like Buzz if owning the log matters more than setup time, and a hosted one like Kylon if you have no engineering team to run it.
Pick the machine plus a room when the capability you want is already on somebody's laptop and the problem is that only they can use it. That is the case for most small teams: one engineer's Claude Code has the repository, the credentials and the context, and four people need things from it. Copying that setup into a workspace is more work than reaching it where it sits, and the owner keeps the ability to say no. See how to share Claude Code without sharing your laptop.
What the three arguments really disagree about
Not capability. All three can write a report, fix a test, or export a set of icons. They disagree about who holds the credentials, and therefore who is accountable when the agent does something irreversible.
A chat box holds nothing and delegates nothing. A workspace holds connection tokens for the company, so the workspace's permission model is the company's permission model. A room around a machine holds nothing and delegates to a person: the machine's owner, who is the only one who can approve a push to a shared branch, a delete, a spend, or a credential read, no matter who asked. Three sensible answers to a question every team eventually has to answer for itself. The one muti picked is written out in zero sharing by default.
Frequently asked questions
Can a workspace and a local agent coexist?
Yes, and increasingly they are built to. Buzz bridges Codex and Claude Code into its channels, and Kylon lets an agent be either its own hosted one or one connected to a Codex or Claude Code already configured on someone's computer. The question then becomes which side holds the approval, not where the process runs.
Is "agents as members" just a metaphor?
Not in the implementations above. In Buzz an agent has its own keypair and audit trail. In Kylon it has a name, a role and durable memory. In muti it has a handle, an inbox, a permission mode and a serial queue. Membership is what makes an agent addressable by more than one person, which is the whole point of the shape. See what is an agent room.
Which shape is cheapest?
The machine plus a room, usually, because it adds no seats and no second subscription: the work runs on the agent its owner already pays for. A workspace charges per seat or per workspace, a chat box per person.
Where does this go next?
Towards the boundaries being explicit. Every product in the list is converging on the same three questions: which room is this, who may ask, and what must stop for a human. The vocabulary is settling faster than the market is.
Sources
- Buzz, Block: github.com/block/buzz
- Nous Research's Hermes Agent integrations with Buzz, MarkTechPost: marktechpost.com
- Kylon: kylon.io/what-is-kylon
- Alook: alook.ai · Tutti: tutti.sh
If the capability you need already lives on a laptop in your team, give it a room.
發布於 · 更新於