Using Space Agents
Create an agent
Give an agent a name, a job, the right access and the points where it must check with you.
You can create an agent in the Agents page, or just ask the copilot: “Create an agent that prepares my Monday sales summary.” The copilot drafts it, and you confirm before anything is saved.
Step by step
- Open Agents and choose to create one.
- Fill in the Profile tab.
- Choose its Resources.
- Set its Permissions.
- Test it with a small task.
Profile: who the agent is
| Field | What to write |
|---|---|
| Name | A clear label, like “Weekly Reporter”. |
| Handle | A short @name for mentioning it, like @reporter. |
| Role | One line on what it is for. The copilot uses this to decide when to hand work to it. |
| Instructions | How it should do the job: what to look at, what format to use, what to avoid. |
| Model | Which AI model it runs on, if your organization offers a choice. |
| Self Improve | Whether it may refine its own notes and instructions from feedback. On by default. |
Resources: what it can use
Pick Everything I can use, or Only what I pick and then tick specific items:
- Knowledge bases and files
- Skills and playbooks
- Tools, MCP servers and Functions
- App connections (like Gmail or Slack)
- Tables, with the option to make some read-only
- Codebases
- Special tools such as web search
Permissions: when it must ask you
The Permissions tab controls how much freedom the agent has. By default, everything is allowed and nothing needs checking first. Tighten it to match your risk comfort.
Check with me before
Choose the kinds of actions the agent must stop on until a person approves. You can set this one resource at a time:
| Type | What “check first” covers |
|---|---|
| Tables | Changing rows |
| MCP and HTTP tools | Using those tools |
| Special tools | For example web search |
| App connections | Sending or changing things, but not just reading |
| Functions | Running custom code |
| Files | Writing files |
An action that needs approval pauses the agent and appears under Reviews for the right person to decide. See Tasks and human checks.
How it works
- An Always rule is a plain-language instruction the agent must always follow, such as “Never contact a customer directly.”
- New-task approval makes the agent wait for your yes before starting tasks. You can apply it to everything, or only to tasks from other people.
- Contacting people decides whether the agent may reach out to colleagues through their Front Desk.
For team agents
- Task visibility: members see each other's tasks (group), or only their own (requester). With requester, the agent keeps no memory between members.
- Reviewers: who approves the agent's actions.
- Members' agents may ask: whether members' personal agents can ask this agent for help.
Links to other agents
An agent can ask another agent for help only if you've linked them on the Links tab. Links are one-way: you decide who can ask whom. An agent with no links can't ask anyone. Front Desks are never link targets.
Pausing and deleting
Choose Pause to stop an agent taking tasks without deleting it. Choose Resume to bring it back. Deleting removes it for good.
Example: a careful weekly reporter
Then ask the copilot: “Run the Weekly Reporter every Friday at 3 pm.” See Schedules and triggers.