An AI coworker is a shared agent that belongs to the workspace. Anyone can mention it, read what it did, and improve its job description. The useful part is not “Claude, but in chat.” The useful part is that the work happens where the team can see it.
What it is
A coworker has three moving parts:
- Purpose. A short job description your team owns.
- Scope. The channels where that job belongs.
- Tools. The CLIs, MCP servers, and APIs it may call.
Everything else is ordinary Claude-style tool use. The difference is ownership: a coworker is shared by the workspace, not trapped in one person’s private session.
What changes
Before a coworker exists, one person asks Claude a question, pastes the answer into chat, and reruns follow-ups for everyone else. The context stays private.
After a coworker exists, the team mentions it in the thread. It reads the channel context, runs the right tool, and replies in place. The next person can scroll up and see the prompt, result, correction, and decision.
Coworker vs chatbot
| Question | Chatbot | AI coworker |
|---|---|---|
| Who owns it? | One user or one app | The workspace |
| What context does it read? | A private session | The channel it joined |
| What can it do? | Usually answer or notify | Run approved tools |
| Who sees the result? | Usually one person | The channel |
Slackbot-style integrations mostly notify. Coworkers act, then leave a visible trail.
Coworker vs ⌘J
Pressing ⌘J opens your personal Claude Code shell inside a channel. A coworker is persistent and shared.
| Press ⌘J | Add a coworker | |
|---|---|---|
| Owner | You | The workspace |
| Tool access | Workspace tools plus local tools | The coworker’s approved tools |
| Lifetime | One session | Persistent teammate |
| Trigger | You press the key | Anyone mentions it |
| Best for | Ad-hoc work | Recurring team work |
For the detailed split, see When to add an AI coworker vs press ⌘J yourself.
Good coworker shapes
Keep the job small enough that the team can tell whether it worked.
- Sales. Reads HubSpot and lists deals at risk this week.
- On-call. Reads PagerDuty and the status page, then summarizes the current incident state.
- Research. Searches arXiv, Notion, and a few feeds, then returns the best two or three sources.
The pattern is the same: narrow purpose, small toolset, one obvious home channel.
Guardrails
A coworker should not act silently, wander outside its tools, or replace a real teammate. If it needs judgment, it should post what it found and mention a person.
Coworkers with three tools usually beat coworkers with thirty. A broad agent picks the wrong path more often and is harder to debug.
When to add one
Add a coworker when the same question keeps coming up across the team and the answer requires the same kind of work. For personal one-offs, press ⌘J. For scheduled or headless work, use the CLI.
Start with Designing a useful AI coworker. For the private counterpart, read Press ⌘J: a shell in every channel.
Further reading
Anthropic’s Managed Agents docs explain the broader tool-use category. For the product-level contrast, read AI coworker vs AI assistant.