An agent is a specialist you build once and reuse everywhere — a researcher, a writer, an operations brain wired into Jira. Here's how to set one up properly, not just spin up a blank name.
Context @researcher
@researcher what did we learn about pricing last week?
Three takeaways from the competitor sweep: usage-based beats seats in this segment, the $20 anchor holds, and annual discounts convert best at 2 months free.
@researcher 3 memories used
Attach an agent with @ — its memory and apps come with it.
1. Create the agent
Start from the Agents page in the sidebar and create a new one. Give it a name that describes the job, not the tool — "Support Triage" tells you more than "Agent 1" ever will.
2. Write real instructions
This is where the agent gets its personality and its job. Be specific: what it's responsible for, how it should handle ambiguous requests, what tone to use. Vague instructions get vague behavior — treat this like onboarding a new hire, not filling in a placeholder field.
3. Pin the apps it needs
On the agent's Config tab, pin the connected apps it should always bring — Jira, email, whatever its job actually touches. You can pin individual actions within an app rather than the whole thing, if you want tighter control over what the agent is allowed to do.
4. Set its memory rules
Every agent keeps its own private memory, separate from your account memory. Decide two things on Config: whether the agent can read your account memory, and whether facts it learns while working can be written back. Keep these off if the agent's work shouldn't bleed into your personal context — a work agent doesn't need to know about your holiday plans, and vice versa.
5. Choose a preferred model (optional)
If the agent's job benefits from a specific model, set it as the agent's preferred model on Config. It brings that model into every chat it's used in, regardless of the chat's default.
6. Bring it into a chat
Once it's set up, type @ in any chat and pick the agent from the list. If you bring in more than one agent, one becomes the boss — it leads the conversation, and when agents disagree on a remembered fact, the boss's memory wins.
Set up once, instructions and apps and memory rules all in place, and the agent is ready to be @-mentioned into any conversation from here on — no re-explaining the job every time.
Try it yourself
Every model, one memory, and agents that work the way you do.