PolyAI Unveils Wren, a Coding Agent for Dialogue Systems

PolyAI aims to make agentic voice AI more human-like than ever. Credit: Getty

Matt Henderson, VP of Research at PolyAI, explains how Wren, an autonomous coding agent, optimises enterprise dialogue agents for better customer service

For years, enterprise contact centres have relied on legacy Interactive Voice Response (IVR) systems.

These platforms force customers through rigid, nested menus which frequently fail to match actual customer intent, resulting in high abandonment rates and profound consumer frustration.

While the initial wave of AI offered a marginal improvement through text-based chatbots and cascaded voice models, these solutions fundamentally lacked the acoustic awareness to handle the natural flow of human conversation, including interruptions, hesitation and background noise.

The solution is PolyAI, an AI company born from the University of Cambridge’s Machine Intelligence Lab in 2017.

Backed by more than US$120m in funding from huge investors, the company has pioneered audio-native conversational systems designed to handle complex, high-stakes enterprise interactions.

Yet the company’s ambition does not stop there. Where most would envision humans as the one improving agentic AI, PolyAI is thinking outside the box with Wren, a coding agent which analyses every agent conversation and proactively improves them, translating into massive enterprise gains.

PolyAI: Growth in numbers

• Estimated valuation of US$500m
• Raised more than US$200m in total funding
• Founded in 2017 in London
• US$86m raised in Series D funding

Hearing the conversation: The audio-native shift

Before an AI can autonomously improve a customer service platform, it must first be capable of actually understanding the caller.

The industry standard for voice AI has historically been a cascaded architecture: an automatic speech recognition (ASR) engine transcribes speech to text, an LLM generates a text reply and a text-to-speech (TTS) engine synthesises the audio.

Yet this method strips the emotional and acoustic context from the call, dropping vital cues before the LLM even processes the input.

To overcome this soulless synthesised audio, PolyAI created Dialog-RSN-1, a frontier dialogue model that perceives and reasons directly over raw audio inputs.

Dialog-RSN-1 combines turn-taking, speech recognition and response generation into a single audio-aware LLM, and can actually catch a caller's hesitation, detect frustration and manage interruptions just as a human agent would.

Matt Henderson, VP of Research at PolyAI, says in a blog: “For the enterprises that build on PolyAI, that means dialogue agents that hear and handle a call the way a great human agent would: catching hesitation, hearing frustration before it escalates and knowing exactly when to speak.”

What sets this model apart is its limit of audio awareness to the input side. Instead, it uses a separate TTS system for generation, allowing enterprises to maintain strict control over brand voice and compliance guardrails while achieving a sub-300-millisecond response latency.

Matt Henderson, VP of Research at PolyAI, says the dialogue agents hear and handle a call the way a great human agent would. Credit: matthen.com

Wren: The agent behind the agents

While deploying a highly capable, audio-native voice agent resolves the immediate friction of customer interaction, the agent cannot just be deployed and forgotten about. Indeed, the long-term challenge for enterprises lies in maintenance and optimisation.

Historically, refining an AI agent required dedicated human engineering teams to manually review a small sample of call logs, identify edge cases and code new conversational pathways. Yet this is time-consuming and costly for the modern company.

Enter Wren, PolyAI’s coding agent built natively into its Agentic Dialog Platform. Functioning as what PolyAI describes as “the agent behind your agents”, Wren autonomously and continuously monitors, refines and optimises customer-facing dialogue agents.

Unlike traditional analytics tools that rely on high-level summaries or small sample sets, Wren analyses 100% of live customer interactions across all channels.

Wren operates through two parallel optimisation motions: self-healing, and self-improving, chiefly.

Arkadiusz Kwapiszewski, Head of Agent OS for PolyAI, explains self-healing and self-improving. Credit: LinkedIn

Arkadiusz Kwapiszewski, Head of Agent OS for PolyAI, explains: “Self-healing catches what's broken in your agent and repairs it, and self-improving finds where something already works but could work better, measured against whatever metrics matter most to you.”

“You decide on the metrics, and Wren will hunt for opportunities to move them. And the impact can be validated through the Agentic Dialog Platform’s A/B testing capabilities.

“First, it engages in “self-healing”, automatically identifying bugs, broken integrations or customer frustrations and generating the code required to repair them.

“Second, it pursues KPI-driven “self-improvement.” Wren proactively searches for opportunities to optimise interactions against North Star business metrics, such as customer satisfaction (CSAT), call containment rates and booking conversions.”

Self-healing catches what's broken in your agent and repairs it, and self-improving finds where something already works but could work better

Arkadiusz Kwapiszewski, Head of Agent OS for PolyAI

Key funding backers of PolyAI

• NVIDIA: NVIDIA provides more than just financial support. It also supplies the crucial hardware infrastructure that enables PolyAI to process speech-to-text, intent understanding and text-to-speech generation in mere milliseconds.
• Khosla Ventures: This venture capital firm has been a steadfast supporter of PolyAI since its earlier funding rounds. Khosla Ventures focuses on early-stage technology companies that have the potential for massive societal and economic impact.
• Hedosophia: Operating as a global investment firm, Hedosophia led PolyAI's US$50m Series C funding round and co-led the subsequent US$86m Series D round. Hedosophia's continued financial support has enabled PolyAI to expand its deployment across hospitality, retail and financial services sectors.