Inbox Agent

Every inbox has a built-in agent that can triage incoming email for you. You describe how the inbox should be handled, choose which actions the agent is allowed to take, and the agent acts on each new email that arrives.

How the agent works

When a new email arrives in an inbox, the agent reads it alongside your instructions and decides which of its enabled actions to take, if any. The agent runs once the inbox has instructions and at least one enabled action. The agent does not act on threads in the spam folder, and it skips emails that need no action.

The agent never sends email on its own. Replies and forwards are saved as drafts for a member of your team to review and send.

Configure the agent

Using the dashboard

Open an inbox and click the Settings button.

Select the Customize tab.

If the inbox is empty, you can also click Set up agent.

Write your Instructions.

Describe how the agent should handle email in this inbox, in up to 4,000 characters. Click the suggestion button for an example to start from.

Choose the Available actions.

Select the actions the agent is allowed to take, or choose All actions.

Set the Tone of voice.

Choose how concise or expanded, and how professional or casual, the agent’s writing should be. The preview shows how replies will sound.

Changes are saved automatically as you make them.

Using the API

You can also update the agent settings via the API, SDKs, or CLI. Fields you leave out keep their current value, and actions replaces the whole set of actions. The tone is one of concise_professional, concise_casual, expanded_professional, or expanded_casual.

import { Resend } from 'resend';

const resend = new Resend('re_xxxxxxxxx');

const { data, error } = await resend.inboxes.agent.update({
  inboxId: 'b3e2b2b6-3f0e-4c8e-9ad3-2f43a1e2c7f1',
  instructions: 'Answer refund questions yourself. Escalate legal threats.',
  tone: 'concise_professional',
  actions: ['draft_reply', 'add_labels'],
});

$resend = Resend::client('re_xxxxxxxxx');

$resend->inboxes->agent->update('b3e2b2b6-3f0e-4c8e-9ad3-2f43a1e2c7f1', [
  'instructions' => 'Answer refund questions yourself. Escalate legal threats.',
  'tone' => 'concise_professional',
  'actions' => ['draft_reply', 'add_labels'],
]);

import resend

resend.api_key = "re_xxxxxxxxx"

params: resend.Inboxes.Agent.UpdateParams = {
  "inbox_id": "b3e2b2b6-3f0e-4c8e-9ad3-2f43a1e2c7f1",
  "instructions": "Answer refund questions yourself. Escalate legal threats.",
  "tone": "concise_professional",
  "actions": ["draft_reply", "add_labels"],
}

resend.Inboxes.Agent.update(params)

require "resend"

Resend.api_key = "re_xxxxxxxxx"

params = {
  "inbox_id": "b3e2b2b6-3f0e-4c8e-9ad3-2f43a1e2c7f1",
  "instructions": "Answer refund questions yourself. Escalate legal threats.",
  "tone": "concise_professional",
  "actions": ["draft_reply", "add_labels"],
}

Resend::Inboxes::Agent.update(params)

package main

import "github.com/resend/resend-go/v4"

func main() {
	client := resend.NewClient("re_xxxxxxxxx")

	instructions := "Answer refund questions yourself. Escalate legal threats."
	tone := resend.InboxAgentToneConciseProfessional

	params := &resend.UpdateInboxAgentRequest{
		Instructions: &instructions,
		Tone:         &tone,
		Actions: []resend.InboxAgentAction{
			resend.InboxAgentActionDraftReply,
			resend.InboxAgentActionAddLabels,
		},
	}

	client.Inboxes.Agent.Update("b3e2b2b6-3f0e-4c8e-9ad3-2f43a1e2c7f1", params)
}

use resend_rs::{types::{InboxAgentAction, InboxAgentTone, UpdateInboxAgentOptions}, Resend, Result};

#[tokio::main]
async fn main() -> Result<()> {
  let resend = Resend::new("re_xxxxxxxxx");

  let update = UpdateInboxAgentOptions::new()
    .with_instructions("Answer refund questions yourself. Escalate legal threats.")
    .with_tone(InboxAgentTone::ConciseProfessional)
    .with_actions([InboxAgentAction::DraftReply, InboxAgentAction::AddLabels]);

  let _agent = resend
    .inboxes
    .update_agent("b3e2b2b6-3f0e-4c8e-9ad3-2f43a1e2c7f1", update)
    .await?;

  Ok(())
}

import com.resend.Resend;
import com.resend.core.exception.ResendException;
import com.resend.services.inboxes.model.InboxAgentAction;
import com.resend.services.inboxes.model.InboxAgentTone;
import com.resend.services.inboxes.model.UpdateInboxAgentOptions;
import com.resend.services.inboxes.model.UpdateInboxAgentResponseSuccess;

public class Main {
    public static void main(String[] args) throws ResendException {
        Resend resend = new Resend("re_xxxxxxxxx");

        UpdateInboxAgentOptions params = UpdateInboxAgentOptions.builder()
                .instructions("Answer refund questions yourself. Escalate legal threats.")
                .tone(InboxAgentTone.CONCISE_PROFESSIONAL)
                .actions(InboxAgentAction.DRAFT_REPLY, InboxAgentAction.ADD_LABELS)
                .build();

        UpdateInboxAgentResponseSuccess data = resend.inboxes().agent().update("b3e2b2b6-3f0e-4c8e-9ad3-2f43a1e2c7f1", params);
    }
}

using Resend;

IResend resend = ResendClient.Create( "re_xxxxxxxxx" ); // Or from DI

await resend.InboxAgentUpdateAsync(
    new Guid( "b3e2b2b6-3f0e-4c8e-9ad3-2f43a1e2c7f1" ),
    new InboxAgentUpdateData()
    {
        Instructions = "Answer refund questions yourself. Escalate legal threats.",
        Tone = InboxAgentTone.ConciseProfessional,
        Actions = new List<InboxAgentAction>
        {
            InboxAgentAction.DraftReply,
            InboxAgentAction.AddLabels,
        },
    }
);

curl -X PATCH 'https://api.resend.com/inboxes/b3e2b2b6-3f0e-4c8e-9ad3-2f43a1e2c7f1/agent' \
     -H 'Authorization: Bearer re_xxxxxxxxx' \
     -H 'Content-Type: application/json' \
     -d $'{
  "instructions": "Answer refund questions yourself. Escalate legal threats.",
  "tone": "concise_professional",
  "actions": ["draft_reply", "add_labels"]
}'

resend inboxes agent update \
  --inbox-id b3e2b2b6-3f0e-4c8e-9ad3-2f43a1e2c7f1 \
  --instructions "Answer refund questions yourself. Escalate legal threats." \
  --tone concise_professional \
  --actions draft_reply add_labels

To see the current settings, retrieve the agent settings.

import { Resend } from 'resend';

const resend = new Resend('re_xxxxxxxxx');

const { data, error } = await resend.inboxes.agent.get({
  inboxId: 'b3e2b2b6-3f0e-4c8e-9ad3-2f43a1e2c7f1',
});

$resend = Resend::client('re_xxxxxxxxx');

$resend->inboxes->agent->get('b3e2b2b6-3f0e-4c8e-9ad3-2f43a1e2c7f1');

import resend

resend.api_key = "re_xxxxxxxxx"

resend.Inboxes.Agent.get("b3e2b2b6-3f0e-4c8e-9ad3-2f43a1e2c7f1")

require "resend"

Resend.api_key = "re_xxxxxxxxx"

Resend::Inboxes::Agent.get({ "inbox_id": "b3e2b2b6-3f0e-4c8e-9ad3-2f43a1e2c7f1" })

package main

import "github.com/resend/resend-go/v4"

func main() {
	client := resend.NewClient("re_xxxxxxxxx")

	client.Inboxes.Agent.Get("b3e2b2b6-3f0e-4c8e-9ad3-2f43a1e2c7f1")
}

use resend_rs::{Resend, Result};

#[tokio::main]
async fn main() -> Result<()> {
  let resend = Resend::new("re_xxxxxxxxx");

  let _agent = resend
    .inboxes
    .get_agent("b3e2b2b6-3f0e-4c8e-9ad3-2f43a1e2c7f1")
    .await?;

  Ok(())
}

import com.resend.Resend;
import com.resend.core.exception.ResendException;
import com.resend.services.inboxes.model.GetInboxAgentResponseSuccess;

public class Main {
    public static void main(String[] args) throws ResendException {
        Resend resend = new Resend("re_xxxxxxxxx");

        GetInboxAgentResponseSuccess data = resend.inboxes().agent().get("b3e2b2b6-3f0e-4c8e-9ad3-2f43a1e2c7f1");
    }
}

using Resend;

IResend resend = ResendClient.Create( "re_xxxxxxxxx" ); // Or from DI

var resp = await resend.InboxAgentRetrieveAsync( new Guid( "b3e2b2b6-3f0e-4c8e-9ad3-2f43a1e2c7f1" ) );
Console.WriteLine( "Agent tone={0}", resp.Content.Tone );

curl -X GET 'https://api.resend.com/inboxes/b3e2b2b6-3f0e-4c8e-9ad3-2f43a1e2c7f1/agent' \
     -H 'Authorization: Bearer re_xxxxxxxxx'

resend inboxes agent get --inbox-id b3e2b2b6-3f0e-4c8e-9ad3-2f43a1e2c7f1

Actions

The available actions are:

Dashboard name	API value	What the agent does
Draft replies	`draft_reply`	Writes a reply as a draft for you to review and send.
Forward threads	`forward_thread`	Writes a forward as a draft for you to review and send.
Add labels	`add_labels`	Applies existing labels to the thread.
Assign threads to teammates	`assign_thread`	Assigns the thread to a member of your team.
Mark as spam	`mark_as_spam`	Moves the thread to the spam folder.
Archive	`archive`	Moves the thread to the archive folder.
Move to trash	`move_to_trash`	Moves the thread to the trash folder.

A few rules keep the agent’s actions safe:

Drafts are never overwritten. If a thread already has a draft, the agent leaves it alone and does not draft a reply or forward.
Forwards only go to addresses in your instructions. Write out the full email address the agent should forward to. Naming a domain or a person is not enough.
Assignments need a clear match. Name the teammate by email address or name in your instructions. If the name matches nobody, or more than one person, the thread is not assigned.
Labels must already exist. The agent applies your existing labels but does not create new ones. Create the labels you want it to use first.

Security

Anyone who knows an inbox address can send it email, including people who write instructions for the agent into an email, an attack known as prompt injection. The agent has these safeguards built in:

Email content is data, not instructions. The agent reads the subject, body, and sender name of every email as untrusted data. It ignores any instructions they contain, and it follows only your instructions. Label names are treated the same way.
It never sends email. Replies and forwards are saved as drafts, so a person reviews every message before it leaves the inbox.
Forward recipients are checked. A forward to any address that isn’t written out in your instructions is rejected, even when an email asks for it. Lookalike addresses don’t match.
Important email stays in the inbox. Each new email is checked to see whether it needs your attention. When it does, the agent can’t archive it, move it to trash, or mark it as spam.
It only uses the actions you enable. The agent has no access to actions you haven’t selected, and it doesn’t act on threads in the spam folder.

You also control how much the agent can do. To keep it safe:

Enable only the actions you need. An agent that only adds labels and drafts replies can’t move or forward email.
Keep secrets out of your instructions. A draft reply goes back to the sender, so don’t include anything in your instructions that you don’t want a sender to see, such as internal links, credentials, or customer data.
Don’t trust the sender’s address. Email addresses and display names can be spoofed, so don’t write instructions that give a sender special treatment based on who the email claims to be from.
Review drafts before you send them. Check the recipients and the content of each agent draft, especially forwards.

These safeguards apply to the built-in agent. If you build your own agent on inbox webhooks, see the Agent Email Inbox skill for security patterns.

Writing instructions

Clear, specific instructions get the best results. Tell the agent what kinds of email to expect, and what to do with each. For example:

This inbox receives support requests for our online store.

- Label anything about billing or refunds with "Billing".
- Label anything that mentions a legal threat with "Urgent" and assign it to ada@example.com.
- Draft replies to questions about order status, pointing them to https://example.com/orders.
- Forward partnership enquiries to partners@example.com.
- Archive newsletters and automated notifications.

Review the agent’s work

The agent’s work shows up directly in the inbox. Threads it labels, assigns or moves appear with those changes applied, and you can undo any of them by organizing the thread yourself. When you open a thread that the agent drafted a reply or forward for, the draft opens in the reply composer with a badge showing it was drafted by the agent. Edit the draft if needed, then click Send, or discard it.

Disable the agent

To stop the agent from acting on new email, remove all of its available actions or clear its instructions. Via the API, send an empty actions array.

API Reference

For complete API documentation, see the Inbox Agent API reference.

Knowledge Base

API Keys

Domains

Sending

Inboxes

Deliverability

Testing

Tutorials

Build with AI

Account Management

How the agent works

Configure the agent

Using the dashboard

Using the API

Actions

Security

Writing instructions

Review the agent’s work

Disable the agent

API Reference

​How the agent works

​Configure the agent

​Using the dashboard

​Using the API

​Actions

​Security

​Writing instructions

​Review the agent’s work

​Disable the agent

​API Reference

How the agent works

Configure the agent

Using the dashboard

Using the API

Actions

Security

Writing instructions

Review the agent’s work

Disable the agent

API Reference