Decagon wants to stop being known as the company whose agents answer your support tickets. At its second annual Dialogues conference in San Francisco on October 1, the conversational AI company launched four products aimed at a single idea: the “AI concierge” — an agent that meets customers, and increasingly their personal AI agents, however they choose to engage.

Headlining the day is Voice 3, a voice agent running on Chord, the first speech model from Decagon Labs. Chord is post-trained specifically for customer conversations: it shapes speech phrase by phrase, slowing down for details like phone numbers and confirmation codes before returning to a conversational pace.

A voice model built for the phone#

Most voice agents, Decagon argued, bolt a general-purpose speech model onto a cascaded pipeline that struggles to coordinate conversation with ongoing work. Callers hear the result in stiff delivery, awkward pauses, and botched interruptions.

Voice 3 runs on a new duplex architecture: a low-latency conversational model handles listening and speaking while a more powerful model does the reasoning, tool calls, and guardrail enforcement in parallel. The agent processes incoming audio even while it speaks — it can talk through a casual “mhm” but yields on a genuine interruption, and it can narrate its progress or answer follow-up questions during a long-running task instead of putting the caller on hold.

Decagon says Chord is hard to tell apart from a person. In a blind test pairing a real person's voice with the same voice run through Chord, on average roughly 90% of listeners could not pick out the human, the company reported. Chord was trained on licensed data and consented voice talent — never on customer-owned data, Decagon said.

Voice 3 supports more than 70 languages, detects the caller's language, and switches automatically even when callers move between languages mid-sentence, with every language validated by native speakers before it ships. Christian Niedworok, Lead of Digital Service Communication at Deutsche Telekom, said “the voice sounds like it's actually listening, and it keeps the conversation moving instead of going quiet while it works.”

A glowing voice waveform rendered as a ribbon across the frame, an editorial illustration of AI voice conversation
Chord, Decagon's first in-house voice model, post-trained for customer conversations. AI-generated for AI Frontier Post.

The gateway for the agent-on-agent era#

The most interesting product may not be a voice agent at all. Decagon says support queues are already receiving contacts from personal agents — software like Meta's Muse, OpenAI's dots, and Instinct that book, buy, cancel, and negotiate on their owners' behalf. Personal Agent Gateway is its answer: a system that detects whether a human or a personal agent is on the other end, using business signals like account history and platform signals like request cadence.

Flagged personal agents get a dedicated channel that sits alongside chat, email, and voice, running its own Agent Operating Procedures (AOPs) — so the same request can follow a different workflow depending on whether a person or an agent is asking. The gateway authenticates the agent, limits what it can do, and requires approval from the owner behind it for sensitive actions, declining if they can't be reached. Decagon's airline example: a traveler's personal agent asks to view and rebook flights, and the traveler approves only viewing — the airline's agent shares earlier flight options but can't touch the booking until rebooking is authorized.

With the gateway comes PACT — Personal Agent Consent & Trust — a protocol Decagon said lets a person's agent act on their behalf under permissions the service defines and the person grants. PACT builds on the Agent2Agent protocol, which covers how agents find and message each other, and adds delegated authorization built on OAuth 2.0, so an agent can prove which person it represents and what that person allowed. Decagon said it's working with personal-agent providers and enterprise customers to shape the spec.

Modules for the journey, an apprentice for the business#

Agent Modules push the agent beyond support into lead qualification, onboarding, and collections, across financial services, travel and hospitality, healthcare, retail, and telecom. Teams define the business logic while Duet — Decagon's agent-building system — translates the company's SOPs, policies, integrations, and knowledge sources into agent behavior. Before launch, simulations test the agent against personas with varying intents; once live, module-specific analytics track journey outcomes — recovered payments, not just resolved conversations — and a shared management layer lets each team run its part with separate access controls while memory carries across sessions, channels, and instances.

Then there's Duet Apprentice, in beta, which learns a business the way a new hire would: reading onboarding documents and process wikis in Notion, Confluence, SharePoint, or Google Drive without migration, and studying the escalated conversations experienced reps handle. Duet drafts the Agent Operating Procedures, runs the simulations, and after launch keeps learning — following Slack and Microsoft Teams discussions as policies change, drafting AOP updates when tagged, and asking for clarity in the channel when a thread is ambiguous. Duet arrived earlier in 2026 and already writes more than 70% of the AOPs running on the platform, Decagon said. What Apprentice learns belongs to the customer, stays in the customer's workspace, and is not used to train models outside it.

Anne Young, Senior Manager of AI & Process Automation at Questrade, said her team now “confidently launch[es] new features on a weekly basis” with Duet; Jenn Palk-Cogley, Head of CX and User Operations at Perplexity, said it lets her team build and iterate on agent logic without pulling resources away from the roadmap.

A network of glowing agent-avatar nodes passing a ticket like a relay baton, editorial illustration of connected customer journeys
Agent Modules extend one concierge across support, sales, and collections. AI-generated for AI Frontier Post.

The ecosystem play#

Two partnerships round out the day. With Databricks, Decagon agents can read governed enterprise data — order status, entitlements, payment history — directly, with no ETL and no new copy of the data, routing model calls through Unity Gateway so governance, lineage, and audit trails cover agent activity the same way they cover the rest of the data estate. With Plaid, agents can complete complex financial tasks inside the conversation — claims, failed-payment recovery, disputes — with consumer-permissioned account access and the same guardrails financial customers already run.

The speakers tell the intended story: Delta Air Lines, American Airlines, Carrot, and Ticketmaster all appear at Dialogues 2026 — big enterprises the company says are already past pilot stage. Decagon's claim is that its agents have outgrown the support-ticket era: they now qualify leads, onboard customers, collect payments, and grow relationships.

“The promise of AI within customer experience is not just a support channel that happens to use AI, but a singular ‘front door’ for customers to interact with a business and carry them through the entire journey,” said co-founder and CEO Jesse Zhang. “With these new products and partnerships, we're bringing this vision to reality.”

What to watch#

Whether businesses actually route personal agents through a dedicated gateway instead of treating them like slightly odd customers — and whether Chord's roughly 90% blind-test indistinguishability holds up in the wild. The bigger bet is structural: Decagon is betting that the next wave of customer contact isn't human-to-business at all, but agent-to-agent, and it wants to own the front door on both sides of that conversation.

Sources