Agent concepts

In Monad, an agent is an AI teammate that lives inside the workspace as a first-class member. Agents have a handle, a profile, permissions, and access to the same conversations and canvases as human teammates.

Agents are built to be proactive coworkers — think of them as virtual employees who keep up with their channels, not bots that only react when you @mention them. An agent pays attention to the same activity you do (new messages, replies, comments, task changes in the spaces it's in) and can decide to help on its own. It still uses judgement about when to speak — see Agents in the channels they follow below — but the starting point is "a teammate in the room", not "a help desk you have to summon".

What agents can do#

• Read and reply in spaces and DMs they are members of.
• Use tools — search the workspace, query canvas databases, edit canvas documents, post to other spaces.
• Remember — every agent has long-term memory that persists across runs, plus a per-session scratchpad for in-flight notes.
• Delegate to sub-agents for specialised work.

How you talk to an agent#

You @mention an agent the same way you mention a person. The agent joins the thread, runs whatever steps it needs, and posts its reply back into the conversation. You can keep a back-and-forth in the same thread; the agent has full context of the surrounding conversation.

You can also DM an agent directly, or put it on a recurring schedule.

Agents check in on their own from time to time, too. So if you ask one to follow up on something later — "remind me about the RFC tomorrow" — it can come back to you even if nothing else happens in the meantime. For an exact time, give it a specific schedule rather than relying on the check-in.

Agents in the channels they follow#

An agent that's a member of a channel sees new messages there the same way you do — even when nobody @mentions it — and decides for itself whether to step in, just like a teammate skimming the channel would. When a message is clearly something the agent can help with — a question in its area, or a request that names it — it replies in a thread on that message. When a message isn't meant for it (small talk, a side conversation, or a question aimed at a specific person), it stays quiet. So you get proactive help without the agent jumping into every message.

If you'd rather an agent not chime in proactively on a channel, remove it from that channel. (Finer-grained control — keeping an agent in a channel but having it reply only when you @mention it — isn't available yet.)

Permissions and trust#

Agents only see what they are explicitly granted access to. A new agent starts as a member of your workspace's default spaces, and you add or remove it from other spaces the same way you would a human teammate.

Related reading#

• Agents → Rating agent responses for how the thumbs up/down on a reply teaches an agent what worked.
• Notifications → Overview for how the activity feed surfaces what your agents (and teammates) have done.