A personal AI agent might be allowed to check a traveler’s reservation without being allowed to change it. For a business receiving that request, the distinction matters more than whether it arrives by phone, chat or a direct agent connection.
That problem sits at the center of Decagon’s Dialogues launch on October 1, 2026. The company introduced four AI products: Voice 3 for customer calls, Personal Agent Gateway for interactions with customers’ AI agents, Agent Modules for industry workflows, and Duet Apprentice for learning from an organization’s internal knowledge.
Together, they mark a shift in Decagon’s pitch from resolving support requests to operating a governed customer-facing concierge. The products are at different stages of maturity. Voice 3 has a described speech stack and company-reported tests. The gateway introduces a new channel and controls, while PACT, the protocol intended to help outside agents use that channel, is still being refined.
Voice 3 Separates the Call From the Work
Voice 3 combines Decagon’s Chord speech model with what the company calls a duplex architecture. It aims to avoid a familiar voice-bot failure: the agent goes silent while checking an account, then interrupts the caller when they offer a brief acknowledgment.
According to Decagon, a low-latency conversational model handles listening, speaking and progress updates while a separate model handles reasoning, tool calls and guardrail enforcement. The two run in parallel. The company says the agent can keep processing incoming audio as it speaks, distinguish a genuine interruption from an “mhm,” and answer a follow-up while a longer task continues.
A caller changing a booking may need to clarify a date while the system checks availability. If the agent can manage that conversation during the lookup, it has a better chance of completing the task without forcing the caller through a rigid sequence. Whether Voice 3 does so reliably in messy, high-stakes calls still requires evidence beyond the launch description.
Decagon also says Voice 3 supports more than 70 languages, detects language changes during a call and offers locale-specific voices. Those are company-stated capabilities, not evidence that every language performs equally well across accents, noisy connections and complex requests.
Chord’s Test Results Need Careful Reading
Chord is Decagon’s speech model, post-trained for customer conversations rather than general narration. The company says it can vary pace and emphasis phrase by phrase, slowing for a confirmation code before returning to a conversational rhythm. Its technical account describes a tokenizer-free diffusion model and a training process designed to preserve pauses and other details of spoken conversation.
