لم تُترجم هذه الصفحة بعد.

The Agents app

This page is for anyone with agents that run for their team. The Agents app is a chat with those agents. You will find an agent, answer it, write to it, and stop or restart it, on a computer or a phone.

The Agents app: your agents and one agent's chat.

Before you start

  • Agents that run for you. The app shows the agents started from the editor, by a rule or by Deploy to…. See Run your instructions as agents.
  • What you see. You see every agent, in each of your organizations, whose instruction you can read. Agents stopped for good aren't shown. Only the people who answer for an agent can answer it, write to it or restart it: see Answer your agents.
  • Sign in. Open instruction.md/agents and choose Sign in if you are asked.

Open the app

Open instruction.md/agents. You can also get there from:

  • a notification that an agent asks you something: it opens that agent's chat;
  • Open in Agents under an agent's conversation in Settings ▸ Deployments;
  • Open the chat with … on an instruction's Agents card, for an agent started this way.

With nothing picked, the page says how many agents you have, how many instructions they follow, and how many run now. Below that:

  • Waiting for you lists what your agents ask you, oldest first. Choose one to open its chat.
  • By instruction shows a card for each instruction, with how many agents follow it, how many run, and how many wait for you.

The app shows your agents from every organization you belong to, together.

Find an agent

Your agents are in the side bar, under Your agents. At its top, choose how to show them:

  • All: the agents waiting for you first, under Waiting for you, then Everyone else.
  • By instruction: each instruction as a group, with each of its agents named by the machine it runs on.

With more than five agents, type in Find an agent to narrow the list.

On a computer, point at an agent to see what it follows: the instruction, the version it runs under Runs, its machine under On, its Rule and its State. From there, choose Open instruction, Edit, or All its agents.

Each agent shows its state:

State What it means
Asks you something It waits for your answer. With more, it says how many, such as Asks you 2 things.
Running It runs.
Not running Its machine says it isn't running now.
Not started yet Its machine hasn't reported it yet.
Waiting for approval It waits for an owner or an admin in Settings ▸ Deployments.
Paused Its rule is paused.
Needs sign-in A tool it uses needs a sign-in on its machine.
Failed It stopped with an error. Its last lines show at the end of its chat.

Subagents

An agent sometimes hands part of its task to a helper, called a subagent. A subagent shows under its agent in the side bar. In the agent's chat, its lines form a block of their own, which you can open and close. Choose Open this subagent to read only its lines.

A subagent works for its agent and answers to it. To tell it something, write to its agent.

Chat with an agent

Pick an agent. The top of its chat links the instruction it follows, with the version it runs, then its machine, its organization and its state.

The chat shows what the agent said and did as it works. Anything it asks you shows as a card, which you answer in place: see Answer a card. Cards show the assistant's reading, as in the editor.

Write to the agent or ask the assistant

At the bottom, choose who you write to:

  • The agent: type in Tell … something… and press Enter. The agent reads it at once, or at its next step. Shift+Enter starts a new line. Until the agent reads it, your message shows not delivered yet.
  • Ask the assistant: ask a question about the agent, such as "What has it done?" or "Is it safe to allow this?". The assistant answers from the agent's conversation and its instruction. The agent doesn't see your question.

Each question counts against your assistant allowance in the agent's organization. If your organization turned the assistant off, the box says To the agent, and you can only write to the agent. See The assistant.

The buttons at the top

Button What it does
Summarize The assistant says in a few lines what the agent did, what it waits on and what went wrong. The summary shows in the chat.
Interrupt The agent stops what it is doing now and waits for you.
Restart Choose Restart now to confirm, or Cancel. The agent starts again from the beginning.
Every agent of this instruction (an icon) Opens the instruction's agents together. See Act on every agent of an instruction.
Deployment settings (an icon) Opens Settings ▸ Deployments of the agent's organization, where its rule is.

The last two always show only their icon. On a phone, the other buttons do too.

Act on every agent of an instruction

Choose an instruction's card, its group in the side bar, or Every agent of this instruction in a chat. The page lists every agent that follows the instruction, with its machine, its version, its rule and its state. Choose one to open its chat.

At the bottom, you can act on all its running agents at once:

  • Type in Tell all … agents something…, then press Enter or the arrow button, Send to all. Each agent reads it at once, or at its next step, and its own conversation keeps it.
  • Interrupt all stops what each agent is doing, and each waits for you.
  • Restart all asks you to confirm with Restart … agents, then starts each agent again from the beginning.

A line then says how many agents it reached, and how many couldn't be reached.

Agents that run in the Claude app

Whoever sets up the machine can choose to run Claude Code agents with Remote Control. Each agent is then a session in the Claude app, named after the instruction it follows. In the Agents app, its chat says Its conversation is in the Claude app., and it has an Open in Claude button. Open it to watch the agent, write to it, approve what it asks and answer its questions, in the Claude app.

For developers: Remote Control explains what it needs.

Sealed agents

In a sealed workspace (beta, not yet independently reviewed), an agent's conversation is end-to-end encrypted. Its chat shows End-to-end encrypted. Choose Unlock and use your key to read and answer it. What you write is sealed in your browser before it is sent. Choose Lock now when you are done. The assistant can't read these conversations, so Summarize and Ask the assistant are off. See Sealed workspaces.

If it doesn't work

What you see Why What to do
No agents run for you yet. No agent of an instruction you can read has been started. Deploy an instruction first: see Run your instructions as agents.
Couldn’t read the agents of … One of your organizations didn't answer. The other organizations' agents still show. Reload the page in a moment.
Your agents couldn’t be read just now instruction.md didn't answer. Choose Try again.
The conversation couldn’t be read: The chat didn't load. Choose try again.
That didn’t go through: try again. Your message or answer didn't reach instruction.md. Send it again.
The assistant couldn’t sum it up just now. The assistant didn't answer. Choose Summarize again later.
Still stuck Something isn't working, or ask us.

Questions

Can I control my agents from my phone?

Yes. Open the Agents app on your phone, or follow the notification when an agent asks you something. You can answer, write, interrupt and restart there, as on a computer. Only the people who answer for an agent can do this.

Why don't I see an agent of my team?

You see the agents whose instruction you can read. An agent stopped for good isn't shown, and neither is one whose instruction isn't shared with you. Ask an owner or an admin to check the agent in Settings ▸ Deployments.

Can I write to a subagent?

No. A subagent works for its agent and answers to it. Write to its agent instead.

Why is there no chat, only Open in Claude?

That agent runs in the Claude app with Remote Control. Its conversation is there, so you open it in Claude.

Last checked against the product:

Was this page helpful? Report a problem