Japanese phrases for instructing an AI agent (and meaning it)

What changed: the agent stopped waiting to be asked

On 23 October 2026, Meta launched its personal AI agent Muse in Japan, the first market outside North America. The company's own announcement describes an assistant you talk to in an app, by text or voice, that plans a task, carries out the steps, and keeps working after you close the app. That is a different kind of software from the chatbot you typed questions into a year ago. A chatbot answers; an agent acts. It opens a page, fills in a field, and comes back with a result.

The part that matters for someone learning Japanese is where that agent is heading. Meta says it is working to connect Muse with services many people in Japan already use, naming the marketplace Mercari, the household-finance app Money Forward ME and the ticket seller Ticket Pia. It also says the agent asks for your approval before it sends a message or buys something, that it runs in its own isolated cloud computer, and that it cannot see your password or card details. Treat all of that as the vendor's description of how the product is designed to behave. It is not a result from independent testing, and I have not tested Muse myself.

So the language problem shifts. You no longer only need Japanese to ask a question. You need Japanese to hand over a job that you still want to control: one clear request, a request to check before acting, a named limit, and a way to take it back. That is exactly what this lesson builds, in seven short steps. By the end you will have a repeatable four-line instruction you can say to an agent, or to a colleague, without sounding vague or reckless.

A concrete way to see the difference: a basic chatbot can tell you how to check whether a train is delayed, but an agent with the right tools can open the timetable, check, and come back to say the 8:12 is running late. The shift is not about what a chatbot is — it is about what you gave permission to touch. Meta's announcement describes exactly that kind of follow-through, and it gives ordinary examples — summarising unread email, managing a calendar — rather than abstract ones. Once a tool can touch your calendar, your inbox or your wallet, the words you use stop being a vocabulary exercise and start being a small piece of supervision.

One request at a time: choosing the right “please”

Japanese gives you several ways to ask someone to do something, and each one signals a different relationship between speaker and listener. 〜てください (“do this, please”) is clear and direct — it tells the other person what you want done, no softening. 〜てもらえますか (“could I have you do this?”) brings the other person's willingness into the sentence: it sounds like you are asking rather than telling. 〜ていただけますか places the speaker below the listener, the form you hear at shop counters and service desks. None of these is a fixed ladder you must climb. They are tools you pick for the moment — a brisk instruction, a considerate request, a humble ask.

An AI agent is software, so nobody's feelings are really at stake. That makes it a safe place to practise the choice. Noticing the shape of a request instead of barking a verb will serve you long after you close the app. A useful default: keep 〜てください for plain setup steps, and reach for 〜てもらえますか when you want the agent — or a person — to hear that you are asking, not commanding.

Play with the three in one instruction. Say the goal first, then one request: 「メールの下書きを作ってください。」 (“Please make a draft of the email.”) Then upgrade it: 「メールの下書きを作ってもらえますか。」 Finally, soften it for a person: 「メールの下書きを作っていただけますか。」 Same job, three temperatures. Notice that only the ending moved; the noun and the verb stayed put. That is the part you want to become automatic.

Each form points a different way. 〜てください is straight — an instruction, and useful when clarity is the whole job. 〜てもらえますか leans on the other person's willingness, so it reads as a request. 〜ていただけますか is the most humble of the three, putting the speaker below the listener — natural at a service counter or in a formal email. A short aside: 〜てくれますか is the plainer cousin of 〜てもらえますか, fine with close friends but a little bare with a stranger. If you remember one safe default, make it 〜てもらえますか.

Ask it to confirm before it acts

The single most useful sentence in this whole lesson is a request to stop and check. In Japanese you build it with 〜前に (“before”) plus 確認 (“checking, reviewing”). 「送る前に確認してください。」 means “Please check before sending.” You can drop this onto almost any action: before sending, before booking, before paying, before deleting. One clause, and you have asked the agent to check before it commits. Whether the agent actually waits for you or just reports what it checked depends on the app, not on the grammar — but saying it clearly is the first step toward staying in control.

Two words will show up again and again in a Japanese agent's interface, and they are worth telling apart. 確認 (かくにん) is checking — looking something over, verifying a fact or a draft. Crucially, the word itself does not mandate a human pause; an agent can 確認 its own logs internally. To ensure you stay in the loop, you must explicitly ask it to wait for your 承認 (approval) before the final act. 承認 (しょうにん) is the word that carries the human sign-off — formally approving so the action may proceed. A screen might ask you to 確認 the details and then 承認 the purchase. Get this straight: 確認 is “review,” 承認 is “green light.” The grammar alone does not enforce a safety control — the app's design does. What the language gives you is the vocabulary to talk about each step clearly, and that clarity is what helps you stay in control.

Here is a clean gated request where you stay in control. 「まず予定を確認してください。それから、私の承認を待ってからチケットを買ってください。」 (“First check the schedule. Then wait for my approval before buying the tickets.”) Three ideas, three verbs, no run-on. まず tells the agent the order. 私の承認を待ってから makes it unambiguous: you are telling the agent to wait for your explicit go-ahead — no room to interpret. This is what a supervised delegation sounds like: the pause is named, and the person who says go is you.

Compare that with a sentence that sounds supervised but is not: 「チケットを買う前に承認をお願いします。」 In this form, 承認をお願いします is the speaker asking the listener for the listener's own approval — you are asking the agent to give its go-ahead, not instructing it to fetch yours. The grammar alone does not tell you who must approve whom. The fix is to name yourself explicitly, as in the first example above. Knowing who is asking whom is the whole game.

Two more forms are worth having ready, and they point in different directions. 〜てもいいですか is a first-person permission question — the understood actor is the speaker, so 「送ってもいいですか」 means “May I send this?” It does not ask the listener to do anything; it asks whether the speaker may act. If you want the agent to stop and wait for you, you already have the tool: 「送る前に、私の承認を待ってください。」 (“Before sending, wait for my approval.”) 〜ないでください asks for restraint — “please don't.” Together the permission and restraint patterns let you say the two things you most want: go ahead, or hold on. 「まだ送らないでください。」 (“Please don't send yet.”) holds the action back until you say otherwise, and it does not pretend the agent itself is asking for anything.

Name the object and the limit

A vague instruction wastes everyone's time, including yours. Japanese is excellent at forcing the details into the open, because the particle を marks what the action touches. 「ホテルを予約して」 leaves the object sitting right there in the sentence: ホテル is what gets booked. When you add a limit, you also add a word your listener is listening for. 予算 (よさん) is budget; 期限 (きげん) is a deadline. 「予算は三万円です。」 is short, plain and impossible to misread: “The budget is thirty thousand yen.”

Then keep the number of moving parts small. 一つずつ means “one at a time,” 一つだけ means “just one.” 「一度に一つずつ進めてください。」 (“Please proceed one at a time.”) is a gift to an agent, because a multi-step task with a checked step after each move is far easier to stop than a single blur of activity. The same habit helps with people: naming one object and one limit sounds calm, while “do whatever seems best” sounds like an invitation to guess.

Here is a limit that also sets a boundary on spending. 「この予算の中で、候補を三つだけ出してもらえますか。予約はまだしないでください。」 (“Within this budget, could you give me just three options? Please don't book yet.”) You named 予算, you capped the output at 三つ, and you closed the door on the action you were not ready for. まだしないでください (“please don't do it yet”) is softer and more precise than a flat “no” — it says later, not never.

Numbers and ranges sharpen a limit fast. 〜まで means “up to,” 〜以内 (いない) means “within,” and 〜だけ means “only.” 「三つまで」 is “up to three,” 「一週間以内に」 is “within one week,” 「これだけ」 is “just this.” Because counters attach to the number, 「三つまで」 reads more naturally to a Japanese listener than a bare “three.” Spend a minute swapping まで, 以内 and だけ onto the same noun and you will feel how each one trims the task a little differently.

Keep a way back

Before you let software do something, know how you would undo it — and what you cannot. Japanese gives you two verbs that look similar and are not. 取り消す (とりけす) means to cancel or revoke something still within your reach: a draft, a pending reservation, an order that has not shipped. 元に戻す (もとにもどす) means to restore or return something to a previous state, like undoing a change you just made. A message you drafted but never sent can be 取り消された; a setting you changed can be 元に戻された. If you mix them up, you end up asking to restore an email that never existed. And if you expect 取り消す to pull back something already delivered — a sent message, a completed payment — the verb will not save you. That limit lives in the service, not in the grammar.

Pair those verbs with 履歴 (りれき), which means a log or history. 操作履歴 (そうさりれき) is your activity history — the list of what the agent did and what it plans to do. Meta says a user can review completed and planned operations, so this is a real screen you should learn to find. And for the emergency stop, one line does the job: 「今の作業をやめてください。」 (「Please stop the current task.」) Short, direct, no room to interpret.

A practical habit: read the history before you approve the next action, not after. In Japanese that sounds like 「操作履歴を確認しました。次に進めてもらえますか。」 (“I checked the activity history. Could you go ahead?”) You confirmed, then you authorised, then you released the next step. The approval stays on your side of the sentence, which is exactly where a careful delegation wants it.

Now the honest limit. 取り消す works for cancellable things — reservations, draft messages, pending orders — but it does not pull a sent email back from someone else's inbox, and it does not refund a completed purchase. A message you already sent to another person is not universally revocable, and no Japanese verb changes that. What the language does is let you stop the action before it crosses the line. 「まだ送らないでください」 (“please don't send yet”) is a gate before the point of no return; 「取り消してください」 (“please cancel”) works only while the thing is still yours to cancel. Learn where that line sits in each service you use, and keep your safety clause on the near side of it.

A pair of handier verbs sits here: 〜ておいてください means “please do this in advance,” and 〜ないでおいてください means “please leave it undone.” 「下書きだけ作っておいてください。」 (“Please just prepare a draft.”) pairs beautifully with 「予約はしないでおいてください。」 (“Please leave the booking undone.”) Notice that 中止 (ちゅうし) is a heavier word — stopping an event or a plan mid-flight — while 取り消す revokes a thing that already exists. Choosing between them keeps your instruction precise about what stage you are actually in.

Make it sound natural, not textbook

Once the grammar works, the next job is to sound like a person. A helpful softener is 〜のですが, which trails off slightly and gives the other side somewhere to respond. Instead of a hard 「これはできません。」 (“This can't be done”), say 「これは少し難しいのですが…」 (“This is a bit difficult, so…”). The sentence leans forward instead of shutting the door. You will hear this constantly in Japanese offices and shops, and it works just as well when you are nudging a tool.

For a light, self-directed question, end with 〜かな. 「これは本当に必要かな。」 is “I wonder if this is really necessary.” It is quiet, thinking-out-loud Japanese, good for questioning your own plan before you send it. And keep the order words close, because an agent, like a busy colleague, reads them first: まず (first), それから (then), そのあと (after that), 最後に (finally). Put まず in front and watch how much easier the whole instruction becomes.

Put these together and a stiff line turns human. Stiff: 「資料を作成してください。」 (“Please create the document.”) Natural: 「まず要点だけまとめてもらえますか。細かい部分はそのあとに相談したいのですが。」 (“Could you first summarise just the key points? I'd like to talk about the details after that.”) Same request, but now it has a first step, a follow-up, and a small opening for the other side. That is the difference between issuing a command and giving a clear, friendly brief.

Two more gentleness tools are easy to overuse and easy to love. 〜ほうがいい gives mild advice — 「確認したほうがいいですよ」 is “you'd better check” — and it softens a suggestion without weakening it. 〜かもしれません adds a maybe — 「まだ終わっていないかもしれません」 is “it might not be finished yet” — which is a graceful way to raise doubt instead of accusing. Used sparingly, both make you sound like a thoughtful person rather than a rulebook, and that is the whole point of this section.

A five-minute practice, and where these came from

Here is the practice. Take one real chore you would happily hand to an agent: summarising a few emails, comparing two train options, or drafting a message. Write it in four lines, in Japanese, and say them out loud once. Line one, the goal, with まず. Line two, one request, ending in 〜てもらえますか. Line three, a pause for checking: 〜前に確認してください. Line four, the real boundary: your explicit approval — 私の承認を待ってください, or a firm まだ送らないでください. The pause is for checking; the boundary is where you say go. Read it back and ask: could a careful stranger follow this without guessing? If yes, you have written an instruction, not a wish.

Then keep the two review words, 確認 and 承認, somewhere you will see them, because they are the two buttons that keep a delegated task under your control. Notice how much of this skill transfers. The same four lines work when you brief a person, because the structure is about clarity and consent, not about machines. Once you have the grammar, what you are really practising is giving an instruction you can stand behind — in any language.

Where this came from, and what it is not. The launch facts and design claims come from Meta's own announcement, Meta、パーソナルAIエージェント「Muse」を日本で提供開始, published 23 October 2026 and read the same week. The planned Mercari, Money Forward ME and Ticket Pia connections are stated in that same announcement. The behaviour described — approval before sending or buying, the isolated cloud computer, the activity history — is Meta's description of intended design; independent testing would be needed to confirm it, and I have not run any. The Japanese phrasing here is standard learner-level grammar, checked for context, but this is not a native-speaker review, and any textbook line should still meet a real person eventually. Prices, features and connectors can change after a launch, so check the current app before you rely on any of it.

Here is a worked example you can copy — one coherent task, four distinct lines. You want the agent to draft a reply to an email but not send it until you approve.

Line one, the goal: 「まず、田中さんからのメールに返信の下書きを作ってもらえますか。」 (First, could you draft a reply to Tanaka-san's email?)
Line two, one clear request: 「内容は、会議の日程を来週の火曜か水曜で提案してください。」 (For the content, please propose the meeting for Tuesday or Wednesday next week.)
Line three, the pause: 「送信する前に、下書きを私に確認してください。」 (Before sending, please show me the draft to check.)
Line four, the limit: 「私が承認するまで、送信はしないでおいてください。」 (Until I approve, please leave it unsent.)

Four lines, four jobs: scope, action, human pause, boundary. Read them back and check one thing: could a careful stranger follow them without guessing? If yes, you have written an instruction, not a wish — and the same four lines will serve you when you brief a colleague.