Using Space Agents
Tasks and human checks
How work is handed to agents, the states a task goes through, and the four places a person can step in.
Giving an agent a task
Any of these starts a task, one unit of work with its own conversation thread:
- Ask in chat: “@reporter, prepare this week's summary.”
- A playbook step that hands work to an agent.
- A schedule or trigger that names the agent.
- A colleague's request through your Front Desk.
A task card appears in the chat. Open it to read the agent's thread and see what it did.
What a task goes through
While running, a task can stop and wait:
| Status | Meaning | What you do |
|---|---|---|
| Queued | Waiting its turn | Nothing; each person runs up to 3 tasks at a time, and a team agent up to 5. |
| Running | The agent is working | Watch, or leave it. |
| Waiting for input | The agent asked you a question | Reply in the task thread. |
| Waiting for review | An action is paused for approval | Approve or decline in Reviews or the thread. |
| Waiting for agent | It's waiting on another agent it asked for help | Nothing. |
| Waiting for approval to start | The task itself needs a yes first | Approve or decline. |
| Done / Failed / Cancelled / Declined | Finished | Read the result, or follow up. |
Talking to an agent mid-task
In the task thread you can:
- Answer a question it asked.
- Request changes by declining a pending review with a reason. The agent adapts and tries again.
- Follow up on a finished task. The agent remembers the conversation.
You can also Stop a task from the chat card.
The four human gates
Every gate becomes a review request, so it shows under Reviews and notifies the right people. The first decision wins.
| Gate | When it triggers | Who decides |
|---|---|---|
| New task | A task needs approval before it starts. Always true for a Front Desk request. | The agent's reviewers, or the task's owner. |
| Action | The agent is about to do something you marked “check with me before”, or it chooses to ask. | The agent's reviewers, or the owner. |
| Resource | It's using a tool, MCP server or skill marked Requires approval. | That resource's approvers, or the person the agent works for. |
| Approval | An approval-needing special tool or app action in a chat or automation. | The owner. |
When named reviewers exist, the person the agent works for can't approve their own agent's action. This keeps a real second pair of eyes on it. If no reviewers are named, the owner approves their own agent's steps.
Safety limits that keep agents tame
- Agents can ask other agents for help only three levels deep, with at most five helpers per task, and never the same agent twice in a chain.
- A task and its helpers share one token budget, so a runaway job can't burn through your allowance.
- A run that is almost out of steps is told so, and answers from what it has instead of failing silently, saying what it didn't finish.
Notifications
You're notified when a task finishes or fails, when an agent needs your input, when a review needs you, when a colleague's request arrives or is answered, and when an agent suggests changes to its own instructions. Items needing you appear first in the bell.