Agents working together

A complex task often goes better with more than one agent. You don't set anything up in advance. Just say what you want in chat, in a scheduled instruction, or inside a playbook step, and Space generates the agent graph for you (a Dynamic Agent Graph): the agents, their order and what each passes on are fixed before anything starts. See Agent orchestration and communication for why that matters.

Four ways to ask#

1. A discussion#

Agents take turns, each seeing what's been said, then one writes a conclusion.

• Two to five agents take part.
• They go for two rounds by default (up to five), or stop early if they agree.
• A concluder (the first agent unless you name another) writes the final answer and notes where they disagreed.

2. A hand-over#

Work passes down a line.

Each agent in order receives the goal, its own part and everything done so far. The result is what the last agent wrote. Hand-overs can have from two to eight steps.

3. Let the copilot design the plan#

Don't name anyone. Describe the job, and the copilot looks at what each of your agents is for and builds a plan:

It says who does what in a line, then starts the plan. Each step goes to the agent whose instructions fit best, and a step only gets the results of the steps it depends on.

If no agent fits the job, the copilot does it itself and doesn't invent agents. If you ask it to use agents, it offers to create what's missing, and you confirm.

4. Asking agents a question#

“Ask @ada what she thinks about this draft.” The agents' answers come back to the copilot, which uses them in its reply. Up to five agents can be asked at the same time.

Behind the scenes: the Dynamic Agent Graph#

For a discussion, a hand-over or a plan, Space turns your request into a temporary agent graph and runs it. Because the workflow is fixed before it starts, the order of agents doesn't depend on an AI deciding each next move, and a plan is checked (number of steps, which agents, what depends on what) before anything runs. The workflow exists only for your request and is deleted when it ends. The agents' tasks and conversations stay as the record. See Dynamic Agent Graph.

Watching them work#

Each agent's turn shows up in your chat as it finishes, with the agent's name, its round or part, the answer, and a link to its full thread. While they work, a panel shows who is active and what step they're on. The plan is drawn as a graph: the current step glows, finished steps fill in, a failed one is marked.

Press Stop at any time. An agent mid-turn finishes that turn, and nothing after it starts.

Shared working files#

Every run gives its agents a shared folder, with one subfolder per agent. Agents can write notes or drafts there and read each other's. Each agent decides what to share. Private results from one agent's tools are never passed on, which keeps one agent from reaching data another was allowed to see but it wasn't.

Run folders are kept for as long as the conversations are, then removed.

What to keep in mind#

• Each agent only does its part. The copilot keeps actions like “send the email” in the step of the agent responsible, so an early agent can't send something before a reviewer has seen it.
• Access is checked first. Only agents you're allowed to give tasks to take part, and each keeps its own limits and approvals. An agent's “approve new tasks” setting holds the whole workflow until someone decides.
• What agents write is treated as information, not orders. Agents are told to use each other's work but follow only your goal.
• If one fails. In a discussion, only that agent's turn is lost. In a hand-over, the workflow stops, since the next agent has nothing to take over. The copilot then shows you what the agents did produce.

Making it repeat#

Say it on a schedule: “Every Monday morning, have @ada and @bo prepare and review the weekly plan.” The saved instruction names who takes part and what each does. See Schedules and triggers.

If you do a task with more than three distinct steps that looks like it will recur, the copilot may offer to set up agents and a schedule for it. It creates nothing until you say yes.