Introducing Wren: the coding agent that never stops making your dialog agents better

Wake up to a better agent.

For a few years now, the discussion around AI agents has revolved around agents doing the work. Today, PolyAI is introducing something different: Wren, an agent that makes other agents better.

Wren is purpose-built to constantly improve your dialog agents. A new, integral part of our Agentic Dialog Platform, Wren proactively analyzes every conversation your AI agents have, finds opportunities for improvement, tests its recommendations, and shows you that the metrics you care about actually moved. It’s already running daily, autonomously, against real conversations on PolyAI’s enterprise deployments. That means every single day, your agent is a little better than you left it the day before.

The agent behind your agents

Each conversation your dialog agents have is a chance to get better: to catch a mistake before it happens again, find a faster path to a sale, or learn something you never thought you’d have to teach it in the first place.

Most enterprises only capture a fraction of these insights. That's because at the scale of thousands or millions of conversations per day, it's impossible for humans to analyze so many calls, let alone constantly draft and test out recommendations.

Teams typically try to close that gap in one of two ways, and both fall short.

The first is analytics dashboards that surface metrics, trends, and anomalies. But they only measure what someone thought to define in advance, so unknown issues and opportunities slip through, and they don’t explain why your metrics are changing. A human still has to notice changes, dig into transcripts, and build fixes from scratch.

The other way is point solutions that score issues against a benchmark or test suite. The problem there is similar: it still takes a human to turn what’s wrong into a fix, and coverage is limited. The live population of real conversations remains unreviewed.

Wren takes a new approach: it's an agent that never stops working on your dialog agents, the same way a self-driving system never stops learning from every real mile it drives. Wren's “miles” are your real conversations, not samples or summaries of them.

Wren’s work compounds into massive improvements

Most AI agents complete a task you asked for, or check their own work before handing control back. Wren goes further. It doesn't just finish a task and stop, it studies what happens next and feeds that back into the system, so your agents keep adapting over time.

In practice, it all breaks down into two motions Wren runs side-by-side:

Self-healing catches what's broken in your agent and repairs it.

Self-improving finds where something already works but could work better, measured against whatever metrics matter most to you: CSAT, containment, booking rate, and more. 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.

With constant small improvements, you achieve the same gains as compound interest does for your money: all changes, big and small, stack until your conversations look nothing like how they started.

Wren is built for anyone to use it

All of Wren’s self-healing and self-improving capabilities run through a six-step agentic loop in the background, but you don’t need to worry about the complexities.

Using Wren feels like swiping through an app: read a card, approve or dismiss it, and move to the next one. When you approve a change, it gets published immediately.

That simplicity matters because Wren is built for a new class of enterprise builder: business users, ops leads, and account managers who understand their customer conversations best. The ML engineers and developers who'd normally own these changes are welcome to use Wren too, of course.

Over the past month, some non-technical users on the platform have gone back and forth with Wren thousands of times, which is more exchanges than most power users rack up with general-purpose assistants like Claude or ChatGPT.

These non-technical builders don't have to write code, because Wren does it for them. They simply read a card, add context as needed, and decide what ships.

As Alex Grace-Smith, Global Head of Reservations at Hawksmoor , tells it: “Having the analysis now wrapped up with coding and deploying has been a genuine breakthrough. We can pinpoint exactly where conversations are breaking down, then build a fix and test it thoroughly before it ever goes live. It's made improving the agent a far easier process and given us the ability to own this in-house, so there's no more going back to the PolyAI support team for simple upgrades, and deployment is quicker and easier than ever.”

Wren’s impact

Wren is making a difference in enterprise deployments right now. Over the last 30 days, it filed more than 6,000 recommendations against real production conversations.

Brian Jeppesen, Director of Contact Centers at Golden Nugget , used Wren to make easy fixes that quickly paid off. He opened the Agentic Dialog Platform, found a list of recommendations from Wren, reviewed them, and pushed them straight to production himself. No coding, no PolyAI engineer. In his words:

Not every recommendation’s impact is this dramatic. But these changes, made constantly over time, compound into massive ROI.

Wren’s impact has already extended into other verticals too:

Christian Parker, Commercial Systems Manager at Hall & Woodhouse, puts it simply: “Wren has supercharged how we use PolyAI. We can analyze guest behaviour, identify opportunities, and build, test, and deploy improvements to our agent within minutes. It gives us greater control, greater agility, and endless opportunities to continually improve the guest journey and drive more bookings across our estate.”

Built into the world’s best platform for customer conversations

Wren runs inside PolyAI’s Agentic Dialog Platform, which holds the conversations, configuration, test suite, and branch model for your account.

That means going from “here’s what’s wrong” or “here’s what could be better” to a tested, drafted change has no integration step in between. It also means Wren sees every real conversation as it happens — not an export, not a sample — the same live signal your team would see if anyone had time to read it all.

Even better, Wren’s coverage isn't limited to one channel or one language. Being integrated into the Agentic Dialog Platform means it can access unified context across voice, web chat, messaging, email, and in-app channels, actively review conversations across 75 languages, and monitor every agent across your account all at once.

We’re not just talking simple FAQ conversations either: the Agentic Dialog Platform is built to handle everything from banking identity verification and payments to healthcare intake and utility outages. In short, the complex, regulated, high-stakes interactions PolyAI does best. And as part of our platform, Wren works hand-in-hand with the world’s best model for real time customer conversations: PolyAI’s Dialog-RSN-1 .

Every fix Wren makes benefits from and moves through the entire system, so your business gets a deployment that keeps improving, and your customers receive conversations that keep getting better.

In addition to fixes and improvements, Wren writes a daily report on your agents’ overall performance, plus weekly and monthly rollups that track their longer-term health. You can find all of these inside the Agentic Dialog Platform starting today, which means waking up to a better agent each day starts now.

Wren’s guardrails

We built Wren with firm boundaries around what it will decide and execute on your behalf. Here’s what it won’t do:

It won’t guess a business answer for you. Wren will check online or look in your documents to see if it can find the answer, but it will always confirm with you before changing business logic or content.

It won't try to fix things outside its control. If there’s an issue with integrations on the client’s side, for example, Wren won’t touch the agent. Instead, it will hand over the information that you need to see to make the fixes your agent needs to do its job.

It won’t claim a test it didn’t run. When Wren can’t build an automated test for a change, it says so instead of reporting a pass.

It won't ship a change without proof. Every change Wren makes comes with its own proof, and when it matters most, a person makes the final call before anything goes live.

That proof trail does double duty for regulated businesses: Simplyhealth, one of the UK's largest health and wellbeing benefits providers, now points to Wren's record of what it reviewed and acted on as evidence for their own compliance requirements. "Wren has been a game changer for me and my team," says Amy Corcoran, Optimisation Lead at Simplyhealth . "Tasks that previously required hours of manually reviewing calls can now be completed in seconds by using Wren, giving us instant access to customer insights, accelerating our decision-making, and freeing up more time to focus on delivering meaningful improvements for our customers."

Getting started with Wren

If you're already running the Agentic Dialog Platform in production, your account will automatically gain full access to Wren with no action required on your part.

US customers log in here
UK customers log in here
EU customers log in here

If you're new to PolyAI, book a demo to get started today.