لم تُترجم هذه الصفحة بعد.

Answer your agents

This page is for the people who answer for agents that run on a machine, with nobody watching. You will answer their questions and permission requests, write to them, and stop or restart them, from the editor or your phone.

Before you start

  • Agents that run for you. They are started from the editor: see Run your instructions as agents.
  • You answer for the agent. The people who answer for an agent are whoever made its rule, whoever deployed it, whoever runs it on their machine, and the organization's owners and admins. Only they can answer it, write to it or restart it.
  • Who is told. Whoever made its rule, deployed it or runs it gets a notification. The owners and admins are told only when there is nobody else, but they can always answer.

When an agent asks you

An agent sometimes needs a person. It wants to know which option you prefer, or it asks before it uses a tool, such as running a command. It then waits for your answer.

Its request arrives as a card. The people told about it (see Before you start) get a notification, which opens the agent's chat in the Agents app. The card also shows in the editor, in Settings ▸ Deployments under Your agents ask you, and in the agent's Conversation.

There are two kinds of card.

  • A question, such as "Which branch should I use?". It has a few choices, each with a line about it. Where it says (you may pick several), you can choose more than one. Or type under Or your own answer.
  • A permission, such as "asks to use Bash". It shows what the agent would do, for example the command it would run or the file it would change.
A permission card with the assistant's reading.

Answer a card

  1. Read the card

    Read what the agent asks, and what it would do. Under it, the assistant explains the request. See What the assistant tells you.

  2. Add a note (optional)

    Type in A note for the agent (optional), such as "No, open a pull request instead." The agent reads it with your answer.

  3. Answer

    For a question, pick a choice or type your own, then choose Send answer.

    For a permission, choose one of these:

    • Allow: the agent may do it this once.
    • Allow for this run: the agent may do it, and won't ask again for the same kind of thing until it restarts.
    • Deny: the agent may not do it. It is told who said no, and reads your note.

The first answer counts. Everyone else who could answer sees who answered, and the card closes for them too. In the conversation, the card becomes one line with the answer, who gave it, and their note.

When nobody answers

Each card shows how long it waits. A permission says, for example, Waits until 14:30, then it is denied. A question says Waits until 14:30, then the agent decides itself. A card waits 30 minutes. A developer can give a rule another time, from 1 minute to 24 hours; the editor doesn't show this choice.

  • A permission nobody answered is denied. The agent goes on without it, or stops and says what it needs.
  • A question nobody answered is left to the agent. It picks the safest option itself and says which one it chose.

Either way, the agent is told that nobody answered. The card then shows nobody answered in time. A card the agent no longer needs, because it ended or restarted, shows no longer needed.

What the assistant tells you

When a card opens, the assistant reads it, the instruction the agent follows, and the agent's recent lines. Under Assistant, it tells you:

  • what the request means, in plain words;
  • how risky it is: Low risk, Some risk or High risk, and why;
  • under Your instruction says, the lines of your instruction that decide it, quoted exactly;
  • the answer it suggests. On a permission, the line says The assistant suggests: Allow, and that button stands out. On a question, the choice it suggests is marked Suggested.

The assistant only suggests. Nothing happens until you press a button. It treats the agent's words as a request to judge, never as orders to follow.

The reading is kept on the card. The next person who opens it reads the same one, and it isn't counted again. If it doesn't appear, choose Explain to ask again.

The reading counts against your assistant allowance in the agent's organization. If your organization turned the assistant off, there is no reading, and everything else works the same. See The assistant for what it sends and what is kept.

Write to an agent, interrupt or restart it

Each agent has a conversation. It shows what the agent said and did, a line at a time, its cards and their answers, and what people wrote to it.

  • In the editor, open Settings ▸ Deployments and choose Conversation on the agent. When it asks you something, the button says Asks you (1).
  • In the Agents app, pick the agent in the side bar. See The Agents app.

From the conversation you can:

  • Write to it. Type in Tell the agent something… and choose Send. The agent reads it at once, or at its next step, and goes on with everything it knew. Until then, your message shows (not delivered yet).
  • Interrupt it. The agent stops what it is doing now and waits for you.
  • Restart it. After you confirm, the agent stops what it is doing and starts again from the beginning.
  • Summarize it. The assistant says in a few lines what the agent has done, what it waits on, and what went wrong.

A conversation keeps its last 500 lines, for 30 days. A card that is still open stays until it is answered or expires.

Answering an agent from a phone.

What leaves the machine

  • A card carries what the agent would do. Before it leaves the machine, anything that looks like a secret, such as a password, a token or a key, is replaced with •••. Long inputs are cut short.
  • The agent's lines are cut short in the same way, with secrets masked.
  • The assistant reads a card only when someone opens it, and reads a summary only when someone asks.

In a sealed workspace (beta, not yet independently reviewed), the conversation is end-to-end encrypted. It shows End-to-end encrypted, and you choose Unlock to read and answer with your key. The assistant can't read these conversations, because it runs on our servers. See Sealed workspaces.

If it doesn't work

What you see Why What to do
(not delivered yet) after your message The agent hasn't read it yet. It reads messages at once, or at its next step. A stopped agent reads nothing. Wait a moment. If the agent shows Not running, ask whoever runs the machine to check it.
No notification when an agent asks Only whoever made its rule, deployed it or runs it is told. Owners and admins are told only when there is nobody else. Check you are one of them: see Before you start. Look in Settings ▸ Deployments under Your agents ask you.
The assistant couldn’t read it just now. The assistant didn't answer, or your assistant allowance is used up. Choose Explain to try again, or answer without it. See The assistant for the allowance.
No reading at all under a card Your organization turned the assistant off, or the conversation is sealed. Answer the card yourself. See The assistant.
That didn’t go through: try again. Your answer or message didn't reach instruction.md. Try again. If someone else answered first, the card shows their answer.
Still stuck Something isn't working, or ask us.

Questions

How do I answer an agent's permission request, and what if nobody answers?

Open the card from your notification, read it, and choose Allow, Allow for this run or Deny. If nobody answers in time, 30 minutes unless the rule says otherwise, the request is denied. See When nobody answers.

What leaves the machine when an agent asks me something? Are secrets sent?

The card carries what the agent would do, such as the command it would run. Anything that looks like a password, a token or a key is replaced with ••• before it leaves the machine. See What leaves the machine.

Who can answer my agent?

Whoever made its rule, whoever deployed it, whoever runs it, and the organization's owners and admins. The first answer counts, and the others see who answered. See Before you start for who gets a notification.

Does the assistant answer for me?

No. It explains the request and suggests an answer, but nothing happens until a person presses a button.

Can I answer from my phone?

Yes. The notification opens the agent's chat in the Agents app, which works on a phone. See The Agents app.

Last checked against the product:

Was this page helpful? Report a problem