Look around. Most AI coworkers fail not because the model is bad but because nobody thought hard about what the coworker was for. The setup flow is fast (give it a name, decide which channels it belongs in, hand it some tools); the design work is the part that matters.
A coworker is shared by the workspace. It shows up in the member list like a teammate. Anyone in the workspace can @mention it. That ownership shape changes the calculus: a coworker isn’t a private tool, it’s a member of the team.
This guide is about the design.
Three things to decide before you set one up
A name. Pick something a teammate would say out loud. sales, oncall, research. Short, lowercase, single word where possible. The name shows up in the channel’s member list, in @mentions, and in the audit trail of who-ran-what. Treat it like you’d treat a teammate’s name.
A purpose. One paragraph. Specific enough that a teammate reading it could repeat the coworker’s job back to you. “Help the sales team” is too broad. “Watch for HubSpot deals that haven’t been touched in 14 days, draft a follow-up email per deal, and post it to #sales for the deal owner to review.” That’s a job.
A toolset. The tools the coworker is allowed to use. Start narrow. Three tools, sharply chosen, will outperform thirty generic ones. The smaller surface forces the agent to pick the right tool more often.
When to add one
The honest answer: when a question keeps coming up in the same channel and the answer is always the same kind of work.
- The sales channel keeps asking which deals are at risk this week.
- The engineering channel keeps asking who’s oncall and what’s paged.
- The design channel keeps asking for hex codes from the brand pack.
Each of those is a coworker. The shape is “the same question, the same place, every few days.”
If a question shows up rarely or unpredictably, you don’t need a coworker. Press ⌘J in the channel and the shell will handle it for the one time it comes up.
What you should expect once it joins
The coworker behaves like a teammate in most ways:
- It shows up in the channel’s member list.
- It has a chip that says “COWORKER” so nobody confuses it with a person.
- It can be @mentioned. It can be muted. It can be removed.
- Its outputs are visible to everyone in the channel. There’s no private mode by design.
When @mentioned, it reads recent channel context, picks a tool, calls the tool, and posts the result. Sometimes it asks a person for input before continuing. That’s deliberate; you want it in the loop with the humans, not silently acting in the background.
Common mistakes
- Too many tools. Already mentioned, worth repeating. Start with three.
- A vague purpose. Vague purposes produce vague answers. Specific purposes produce useful ones.
- Treating it like ChatGPT. It’s not a general assistant. It acts when @mentioned, then goes quiet. If you want a general-purpose chat partner, use Claude directly. Coworkers are job-shaped agents.
Going deeper
The concept of an agent that lives in a channel is in the broader family Anthropic describes in their Managed Agents docs. That’s the right place for the upstream picture of how the model, the tools, and the run-loop fit together.
For the conceptual framing of what a coworker is and isn’t, see What is an AI coworker? and AI coworker vs AI assistant. For the rubric on when to spin one up versus pressing ⌘J yourself, see When to add an AI coworker vs press ⌘J yourself.
Once you’ve designed the coworker, the actual setup happens inside the Ano desktop app, free to download during the open beta.