FeaturesNews

Meet Agents: assistants with their own memory, tools and routines

Set one up once, give it a job, and it works from a chat that feels like texting. It remembers what you allow, runs on a schedule, and messages you when it has something to say.

Scroll down

A chat is a good way to ask a question. It is a poor way to keep an eye on things. You do not want to open a conversation every morning to ask what happened overnight, or remember to ask the same thing again on Friday.

Agents are for that. An Agent is an assistant you set up once, with a job, instructions of its own, and a chat that works like texting. It can remember what matters, run on a schedule, and message you when it has something to say.

An Agent named Scout in its Home chat giving a morning summary of issues and pull requests, asking whether to draft review notes, and then listing three notes after the Draft the notes button is chosen.
An Agent's Home chat. The project, issues and pull requests shown are samples.

Set one up in a minute

Choose New agent and give it a name, a purpose of a sentence or two, and instructions if you want it to work a particular way. The instructions can be long, up to 8,000 characters. If you would rather not write them from scratch, the app drafts a starting brief from the purpose while you watch, and you can edit all of it.

An Agent also gets a face. Pick a shape, a colour and an expression, from calm to curious to sleepy, or upload your own picture.

What an Agent has
  • An Agent
  • Instructions. Its job and how to do it
  • Memory. Account, read only, or off
  • Skills and Plugins. Only what you turn on
  • Routines. Run on a schedule
  • Home chat. Replies like texts

Its Home is a text thread

Every Agent has a Home chat, and it behaves like a conversation with a person. Replies arrive as short messages, one at a time, not as a wall of text. An Agent can ask you a question and offer buttons to answer it, ask you to sign in to a connector it needs, and attach a chart.

The screenshot above is an Agent named Scout giving a morning summary of a project's issues and pull requests, then asking whether you want review notes drafted for the one that is waiting. You tap Draft the notes, and it does.

It remembers what you tell it, and only as much as you allow

Each Agent has its own memory setting. Use account settings follows what you chose for the rest of Chat. Read only lets it use what has been saved without adding new facts. Off gives it no memory at all. It is a switch on the Agent, not something you have to reconsider every time you talk to it.

Give it only the tools it needs

Each Agent has its own Skills and Plugins screen, with search, an All, On and Off filter, and sorting. By default an Agent can use everything you have turned on. Switch things off and it has less to work with. This is worth doing for an Agent that touches anything you care about, such as a trading plugin, which asks you to confirm the risk first.

Routines that run without you

An Agent can run on a schedule. Create a routine under Scheduled Tasks, choose the Agent to run it, and the result lands in that Agent's Home chat. The chat shows when the routine starts and how long it ran, and the result arrives as a message like any other. You can set it to run once or to repeat.

Switch Agents in the middle of a chat

You do not have to start over to change who you are talking to. Pick another Agent in the middle of an ordinary chat, and the next reply uses that Agent's instructions, memory, skills and connectors. An Agent's own Home chat always stays with it.

On your phone

On iPhone, the Agents list shows each Agent's newest reply, and says Working… while it works or Waiting on you when it needs an answer. A row's menu offers Rename, Customize, Pin and Delete, and you can swipe a row to delete it and tap Undo if you change your mind.

When an Agent replies and none of your devices is showing the conversation, you get a notification with the first line of the reply. Notifications are part of Pro and Max, and each Agent has its own switch.

Agents can also use your computer

An Agent can work in your browser or in an app on your computer, if you allow it. That has its own rules, described in Let an agent use your computer.