Give voice agents an email inbox

Voice agents close the call and then forget. The customer still needs a recap, a link, or a way to send a PDF that does not fit in a spoken sentence. Phone and WebRTC are the audio plane. Email is the durable plane. Give the voice role its own Agent Inbox so people can reply when they are no longer on the line.

Use any voice runtime you already have — LiveKit is a common example of an audio stack, not an exclusive partnership. Neuro OS runs the role that may send mail only after Ask, and that forwards a thread when a human must take the next call or the next reply.

After the call: draft a recap

When the session ends, the voice agent (or a sibling role) should produce a structured recap: who, when, decisions, open questions, promised files. That recap becomes a draft email to the participants who opted in. It does not send itself.

Ask shows the draft, the recording policy reminder, and the thread it will start. A person edits anything the model invented — especially prices, dates, and legal phrasing — then sends. If nobody is available, the draft waits. A late accurate recap beats an instant wrong one.

Create a stable inbox per voice role or per queue:

from agentinbox import AgentInbox
client = AgentInbox()
inbox = client.inboxes.create(username="voice-support", domain="agentinbox.space")

Do not use the on-call engineer’s personal mailbox as the From.

Receive documents by email

Customers will say “I’ll send the contract” and then actually send it. The voice agent cannot receive a PDF on the RTP stream. Tell them, during the call, the role’s address. Inbound mail lands on a thread keyed to the customer. The role extracts attachments, stores them in the project workspace, and drafts a confirmation.

Confirmation mail is still Ask. Auto-acking “we got your file” is how you confirm the wrong file to the wrong party. Match on the address they used and the open ticket, not on a similar name.

Forward to a human

When the call requires a specialist, or the inbound mail is angry, or the attachment is a legal notice, forward the thread into Neuro OS. The human sees the recap, the transcript link if policy allows, and the mail. They reply from the same thread so the customer is not bounced to a new address.

The voice runtime stays responsible for audio: joining rooms, streaming, barge-in. Agent Inbox stays responsible for mailbox: threads, attachments, sends. Do not bolt SMTP onto the voice worker. Do not put trunk credentials in a skill file.

Channel boundaries

Voice is real-time and unforgiving. Email is asynchronous and auditable. The agent should not read a long thread aloud unless the caller asks. It should not send mail in the middle of a live call without Ask — there is no reviewer on the RTP path. Queue the draft for after the hangup.

CRM writes from the call (disposition, next step) default to Ask as well. The recap email and the CRM patch can share one approval card so the reviewer sees both.

Start with recap-only. Add inbound documents second. Add human forwarding third. Measure how often recaps are edited. That edit rate is the quality signal, not how fast the TTS said goodbye.

Audio plane and mailbox plane, connected by a role with gates: that is the integration. No exclusive vendor story required.

Agent Inbox gives each role a mailbox people can reply to, with forwarding into Neuro OS when a human must see the thread. Outbound mail defaults to Ask. Run the role on Neuro OS. To scope the first inbox, get started.

Explore Neuro OS →