Build an agent for your team
A team agent does one kind of job for everyone on a team, such as “Juniper Research Desk”. It works the same way for everyone, and people can reach it in Context, by email or in Slack.
First check that you can see New agent in the Agents panel. On most organizations Members need an admin to allow agent creation for their role.
Then describe the job in one sentence. “Answers product questions from the sales team using our product docs” gives you something to test against. “Helps with stuff” doesn’t. Keep each agent to one job, because a focused agent is easier to test and easier for people to trust.
Create the agent
- Go to Agents > Starting points, find one close to your job, and click Use. To start blank, click New agent instead.
- Give it a name people will recognize and a one-line description of what it does.
- Fill in Setup:
- Model: leave the default unless you have a reason to change it.
- Instructions: what the agent does and how. Type @ to link a skill, file, connector, agent or person.
- Skills: your team’s skills, such as your report format.
- Files: reference files it should always have.
- Connectors: only the apps it needs. Each one is somewhere it could take action.
- On Access, choose which workspaces it works in, and which people or user groups can use it. If people later say they can’t find the agent, check this tab first.
- On Permissions, decide which actions it may take on its own and which need approval. New agents start on Full access. Decide what agents can do explains the choices.
- Click Create agent.
Start from the team’s workspace
You can also start in the workspace the agent is for. When a workspace has no agent, its home shows “No agent in this workspace yet” with Create an agent. If you already have an agent that could work there, Add your agents brings it in instead. Create an agent appears only if you’re allowed to create agents.
The agent dialog has three tabs: Setup, Access and Permissions. Only the name is required. Give it a name and a short description, then write its job under Instructions in plain words, for example “Write a one-page account brief for each client company: what they do, recent news, and three questions for the next meeting.” Under Skills go the procedures it loads when a task calls for one, and under Files the files it reads before every task. A file you upload there is saved to your Drive first.
Add it to workspaces
On Access > Add to workspaces, the workspace you started from is already picked, if an agent can be added to it and you can add members to it. Remove it or add more under Add to a workspace….
Personal workspaces are never listed, because agents can’t be added to them. On organizations that don’t use the newer permissions system, private workspaces aren’t listed either.
Adding a workspace doesn’t share the agent with that workspace’s members. That’s the next step.
Make it available to the team
On Access > Make available, the agent shows “Members of [workspace] cannot use it yet.” Click Share with them as and pick a role:
| Role | What they can do |
|---|---|
| Viewer | See the agent, but not message it. |
| Operator | Message it and use it in tasks. |
| Admin | Also change its setup, memory and access, and delete it. On the newer permissions system, Admin can’t delete it. |
Pick Operator so everyone in the workspace can use it. To add named people or user groups, type in Add people or user groups…; they start as Operator. You are the agent’s Owner, and ownership can’t be shared.
On the newer permissions system. An agent is shared with one person at a time. Sharing with a workspace’s members or a user group is refused with “On this organization an agent is shared with one person at a time.”, so add each person by name in Add people or user groups… and pick Viewer, Operator or Admin. Only the agent’s owner or an organization admin can delete it.
Choose a permissions preset
On Permissions, open the Preset menu. New agents start on Full access. For a team agent, pick Ask for changes. With it, the agent:
| Group | Runs on its own | Asks first |
|---|---|---|
| On its own computer | Running terminal commands, editing its own files, opening webpages, using skills | |
| Your files in Drive | Reading your Drive files | Editing your Drive files |
| In your workspace | Moving, renaming, and sharing files; creating and editing skills; creating and editing documents; changing workspace settings and members; creating or changing agents; building or changing applets; adding or removing connected apps | |
| Apps you’ve connected | Using connected apps |
“Its own files” are the agent’s working files on its own computer, not your Drive. The other presets and how to change single rows are in Decide what agents can do.
What “Ask first” and “Don’t allow” mean
Ask first sends you an approval. The table’s footer sums up the choices as once, this task or always: on the approval they read Allow once, Allow all … until this task is done and Always allow …, and the last two appear only when they’re offered for that action. Don’t allow is final: the agent can’t ask its way past it. See Answer approvals.
When it’s ready
Click Create agent. When nothing is left to do, the dialog says “[Agent] is ready” and “Everything is in place.” Click Start a task to open a task with the agent, or Done.
If something is still open, the dialog says so instead. For example, if the agent isn’t in any workspace yet, the @ picker won’t show it.
Back in the workspace, the agent is listed under Direct messages while that list is still empty, on organizations where channels and direct messages are turned on. People with Operator access can message it there. In the task box, the agent is already picked for you because it’s your newest agent. Teammates pick it once in the task box, and after that it’s their default in this workspace.
Write the instructions
Say what to do in which situation, and what not to do. “Be helpful, accurate and fast” gives the agent nothing to act on. For example:
You are the Juniper Research Desk. You answer questions from the sales team about our products. Always use the files in @Product Docs first, and say which file you used. If the answer is not in our docs, say so; don’t guess. Keep answers under 150 words unless asked for more. Use the @Juniper tone skill.
Keep house style and repeatable steps in skills rather than in the instructions. Skills can be shared with other agents and improved over time.
Test before you share
Send the agent several real jobs from a workspace, and fix the instructions after each one. Do this before the team starts using it.
For the first month after it goes live, check the Activity tab each week. It shows what people are really asking.
Reach it outside Context
Open the agent’s Platforms tab:
- Email: the address people can write to. It says “Not set up” until the agent has one.
- Slack: give it its own Slack app. Click Create app, then Add to Slack. An admin must first turn on the Slack app for the organization.
- Microsoft Teams: available once an admin connects your Microsoft tenant.
- Phone, SMS and iMessage are shown as “Coming soon”.
On newer organizations. With the newer permissions system, Members can create agents by default. An agent also works with its own access, not the access of the person who asked it.