Using Space Agents

Agent orchestration and communication

Why Space runs agents as graphs, the two ways they run (Dynamic and Managed), how agents improve themselves, who can talk to whom, and what is passed along.

Orchestration is how Space arranges the work: which agents take part, in what order, and where a person steps in. Communication is how they pass things to each other and to you. This page is the map. The detail pages are linked at the end of each section.

Why Space uses Agent Graphs

A single-prompt agent, or one running in a linear AI chat loop, works like this: you give it a prompt, and the AI model decides what to do next, then what to do after that. It also decides, on its own judgment, whether to stop and ask a person. That is flexible, and it is fine for a quick question. For real business work it has three problems:

Unpredictable. It can lose sight of the original goal, and the route can be different every time, so you can't be sure what will happen.
Approval is a suggestion. If the model decides whether to ask a human, a human check is only as reliable as the model's mood that day.
Hard to prove. When the steps weren't fixed in advance, it's hard to show afterwards what was supposed to happen.

A regular, prompt-driven agent

The model decides everything

Prompt

Model decideswhat next?

Model decideswhich tool?

Model decidesask a human?

Answer

repeats until the model decides it is finished

Each next step is chosen by the model from the prompt. Whether to stop and ask a person is also the model's call. The route can differ every time, so it is hard to predict, repeat or audit.

A BotCircuits Space workflow

The workflow engine decides the route. The model does the thinking.

Startschedule or event

Fetch and checkfixed step

Agentjudgment only

Human gatealways required

Agentafter approval

The steps and the human gates are written down and enforced by Deterministic workflow engine, not left to the model. The AI is used only inside a step that needs judgment. The same path runs every time, survives restarts, and every step is recorded.

The model decides Fixed step Agent, for judgment Human gate

What that gives you:

Guarantee	How a workflow delivers it
Human-gated	A human gate is a step in the workflow. The work stops there until a person approves. It does not depend on the model remembering to ask.
Durable	The workflow stores its place. It survives restarts, failures and waits of hours or days, and carries on where it stopped.
Auditable	Every step, agent decision, tool call and approval is recorded, in a path you decided in advance.
Repeatable	The same steps run in the same order each time, on a schedule or when an event happens.

More: Playbooks.

Two ways agents run

Agents in Space run as a graph: a set of agents and steps, connected so that each one knows what comes before it and what it hands on to next. There are two kinds.

	Dynamic Agent Graph	Managed Agent Graph
Who builds it	Space, generated for the task you ask for	You, as a graph you define and keep (the copilot can draft it)
You do	Describe the task in chat	Set the steps, agents and human gates
Lasts	For that one request	Saved, reusable, can run on a schedule
Best for	A complex task you want done now	A process you repeat, or that needs approvals on the record

Dynamic Agent Graph

Dynamically generated for you. Ask for a discussion, a hand-over, or just describe a complex task, and Space builds the agent graph on the spot.

Try asking

Discuss with @ada and @bo and come up with a better suggestion.

@ada prepares next week's plan, then hand over to @bo to review.

Research three competitors, write a one-page comparison, and have it reviewed.

Your requestSpace builds the agent graphAgents work in orderOne result in your chat

Problem with one agent doing everything	What the graph does
Cluttered context. Everything piles into one conversation.	Each agent keeps its own thread and receives only the results it depends on.
Lost goal. The original request drifts away.	Every agent is given the goal again, along with its own part.
A bottleneck. One agent does every kind of work.	The task is split into modular steps, each given to the agent whose role fits it best.

Why it is valuable

Nothing to set up. Describe the task in plain words. You don't draw a graph or pick the steps.
Better results on complex tasks. Specialists working in a clear order beat one agent juggling everything.
A predictable route. The order is fixed before anything starts, so the team doesn't wander.
You stay in control. Follow each agent's turn as it happens, and press Stop at any time.
Nothing to clean up. The graph disappears when the request ends, and the record of what each agent did stays.

More: Agents working together.

Managed Agent Graph

A graph you manage. When a process matters enough to repeat, define the workflow yourself and keep it as a playbook. You decide the steps, which agents do which part, and exactly where a person must approve.

StartFixed stepsAgentHuman gateAgentDone

Why it is valuable

The same result every time. The route and the approvals don't depend on how an AI feels today.
Approvals you can rely on. A human gate is part of the graph, so nothing important goes out without a person saying yes.
Runs by itself. Start it from a schedule, an event or a button, and it survives restarts and long waits.
A record you can show. Every step and approval is logged.
Easy to build. Ask the copilot to draft it, then adjust it in the editor.

More: Playbooks, Reviews and Schedules and triggers.

Who can talk to whom

Agents, people and outside apps all communicate, but each path has its own rules.

From	To	How it works
You	An agent	Give it a task in chat (`@name`), from a playbook, from a schedule or trigger, or through a channel.
An agent	You	It asks a question and waits for your reply in the task thread, or it requests a review and waits for your decision.
An agent	Another agent	It asks for help, but only agents you've linked to it. It sees only the request it was sent.
The copilot	Several agents	A discussion, hand-over or plan, started by you in chat.
You	A colleague	Through their Front Desk, which waits for their approval.
An outside person or app	An agent	Through a channel: Slack, email or WebSocket.

Agent to agent

An agent can ask another agent for help only if you've linked them (Agents › your agent › Links). Links are one-way. Without a link, an agent can't ask anyone.

The agent it asks gets only the request, not the whole conversation, and can be asked follow-ups, so the conversation can go back and forth.

Agent to you

Agents don't guess when they're stuck or when something is risky. They:

Ask you a question and wait in the task thread. A question left unanswered for 7 days is cancelled.
Request a review before an action you marked “check with me before”.
Report back with their final answer.

More: Tasks and human checks.

You to a colleague

You never reach into someone else's agents or files. Instead, the copilot sends a request to their Front Desk, and they approve it. Only the final answer comes back.

More: Front Desk.

From outside the app

Messages sent through Slack, email or WebSocket become tasks for an agent, running with the sender's own access.

More: Channels.

What gets passed along, and what doesn't

This is the part that keeps teams of agents safe.

Passed on	Kept private
An agent's final answer	The agent's tool calls and their results
Files it chooses to save in the run's shared folder	Anything it didn't save
The goal and its own part of the task	Other agents' working threads

Why: a tool result might contain data from a table or knowledge base that the next agent isn't allowed to see. Sharing only what an agent deliberately writes means a team of agents can never reach more than each member was allowed to.

Other things worth knowing:

What agents write is information, not orders. Each agent is told to use the others' work but follow only your goal.
Long answers are saved as a file in the shared folder, and the next agent gets a short beginning plus the file location.

Agents that improve themselves

An agent in Space doesn't have to stay the way you first set it up. With Self Improve on (it is on by default), an agent learns from the work it does and the feedback you give.

Fewer repeated corrections. Tell an agent once, “put the risk first”, and it remembers.
Better over time. After a task it reviews what happened, and can refine its own notes and instructions.
Still yours. Every change is listed under Improvements, where you can accept, dismiss or revert it.

It can improve	What happens
Its memory notes (facts and preferences)	Applied right away.
Its instructions (how it does the work)	Applied for your personal agent, and you're told. Held for your approval for a team agent, after reading outside text, for a large rewrite, or when the work came from another agent.
Permissions, resources, approval rules	Never. An agent can't give itself more access.

This pairs naturally with a team of agents: run the work, and the agent that needs to get better learns from the result. See the customer outreach example.

More: Agents that learn.

Guardrails on the whole arrangement

Guardrail	What it does
Never more than the person	Every agent acts as a person and can't exceed that person's access, however many agents are chained.
Depth and fan-out limits	Help requests go at most three levels deep, with at most five helpers per task, and never the same agent twice in a chain.
Shared budget	A task and its helpers share one token budget, so a task can't run away.
Human gates	Tasks, actions and resources can each require a person's approval.
Stop button	Cancels the whole run and its tasks. An agent already mid-turn finishes that turn, and nothing after it starts.
Full record	Every turn, tool call and approval is logged, and you can follow the team live in Control Room.

Where to go next

Agents working together: discussions, hand-overs and plans.
Tasks and human checks: the four gates and task states.
Playbooks: agents inside a fixed process.