Skip to main content
An agent is a configured AI teammate. You give it a purpose, decide when it wakes up, choose which tools it can use and which data it can reach, and then it works on its own, producing sessions you review afterwards. Chats are how you answer a question once. Agents are how your team stops answering the same question by hand.

​
Agents or chats?

Both run on the same tools and the same underlying AI. The difference is who starts them, and whether the setup is worth keeping. Reach for an agent when the work is repeated, has a shape, and someone keeps doing it by hand. A weekly report. Every partner lead list that lands in a ticket. Every question in #marketing-ops that has the same answer.
Start in a chat. Once you’ve done a task twice the same way, that’s an agent.

​
What you configure

An agent’s name, description, and icon are set from its header, along with Pause, which stops its triggers waking it. Everything else lives on four tabs, and they build on each other. Every one narrows what the agent can do. Nothing widens it.
Two things live here.Triggers: when the agent wakes up. An agent with no triggers only runs when you start a session yourself, which is how you test one before turning it loose.Context: the instructions it follows in every session, and documents you upload for it to reference. Instructions are prose the agent always has in mind; documents are reference material it reads when it needs to, such as lookup tables, field mappings, style guides, long specifications. Put stable guidance in the instructions and bulk data in documents.
Slack and Linear each need to be connected for your workspace before an agent can be woken by them. Set that up in Workspace settings, under AI. See Slack and Linear.
Every tool can be set to one of three modes for this agent:A disabled tool isn’t available to the agent at all, so it can’t decide to use it anyway.
An agent has its own permissions, separate from yours, the same way an API key does. It can only reach what you’ve granted it.Tools and access are enforced separately. An agent allowed to use a tool still can’t use it on data it has no permission for.

​
Sessions

Every time an agent runs, it produces a session, one conversation, with the full history of what it did. Sessions are listed on the agent’s Overview tab. A session works in turns: a message comes in, the agent thinks and takes actions, then replies. A session can span many turns over days, because a Linear issue or a Slack thread stays open. Two behaviors are worth knowing:
New messages take priority. If you message an agent while it’s still working, it stops what it was doing and starts on your new message instead. The newest instruction wins, so if you want it to finish first, wait for its reply.
Approvals wait as long as they need to. When an agent hits an action that needs approval, it stops and waits. There’s no timeout. Answer it an hour later or the next morning and it picks up where it left off.While it’s waiting, it isn’t listening. Messages sent to a session that’s holding an approval are ignored. Answer the approval first, then send your message.

​
Agents and approvals

Agents use the same approval system as chats, but you decide per agent which actions pause. This is the main control you have over an autonomous agent, and it’s worth setting deliberately:
  • Leave an action on Requires approval and a person confirms every time. Good for anything that writes to your CRM or messages a customer.
  • Set it to Automatic and the agent proceeds unattended. Good for reads, and for writes you’ve watched enough times to trust.
  • Disable it and the agent will never do it, no matter how it’s asked.
An approval prompt reaches whoever is watching the agent’s session. If your agent runs unattended overnight, an approval simply waits until morning.
An agent set to run fully automatically will act without anyone watching. Start new agents with writes on Requires approval, watch a few real runs, and only then loosen the ones you’re confident in.

​
Where agents run

​
More resources

Triggers

The four ways an agent can be woken up, and how to scope each one.

Slack

Connect Slack, and how an agent behaves in channels and threads.

Linear

Connect Linear, filter which issues an agent picks up, and what it can write back.

Approvals

Which actions pause for confirmation, and how to review them.