What I learned building voice agents at Decagon

I've always been drawn to agentic voice. The pace of the competition and the industry, the complexity of all the moving components, the massive impact on the business, and, if I'm being honest with myself, because I love to talk. Build an agent that talks back to me? Sign me up.

When I got the chance to focus on it, the Decagon Agent Development team was already rolling out agents on our other channels almost daily. Chat, email, SMS, live and resolving real conversations like a machine. So voice should be no different, right? Not quite.

Voice is magic for the end user. No more hold music to cancel a reservation. No more pressing five numbers to check on a prescription. No more repeating yourself to four people before someone finally helps. But to the person on the line, that magic is invisible, and to the team building it, the pain didn't disappear. It transferred to us.

In the early days, we sank hundreds of hours a week into a single voice agent. Everything I knew from text channels was 10x harder here. The architecture was more complex, the information collection more nuanced, and the bar for a good experience was unforgiving. So I did what the work demanded.

To learn, I followed our voice engineers around the office until they explained how TTS works, then begged them to join customer calls to explain it live.

To build, we ran dozens of AOP and conversationality experiments, hunting for the one phrasing tweak that changed everything.

To test, we bribed every Decagon employee we could into one room to replicate a call center (yes, ironic) and stress-tested every scenario we could think of. At one point I was doing my best granny accent to see how transcription held up. Voice acting may be my next calling.

To iterate, I combed transcripts line by line and ran latency analysis that would make you pull your hair out, hunting for milliseconds to shave.

Then we shipped proactive outbound calling, and the difficulty curve went vertical. Outbound is longer, higher stakes (medication refills, collections), and the branching logic is gnarly. A bad agent interrupting your day? You hang up. An intelligent, empathetic one calling to make your day easier? Five stars. The gap between those two is small in the build and massive in the outcome.

None of it was scalable. We were shipping high-CSAT, high-deflection agents in chat and email like clockwork, but voice was different. And voice was about to be everywhere. We called what we saw coming the voice tsunami.

In December 2025, a few of us sat down to figure out how to ride it. The question was simple to ask and hard to answer: what does it actually take to build voice agents at scale? We tracked every piece of work that went into a voice launch, backfilling tickets for things we'd already shipped so we had a real picture of the lift. My teammate Kevin Liu ran the analysis that showed Agent Development and Voice Product/Eng exactly where to focus.

Then we built the system:

• Tribal knowledge into playbooks. Everything in my head, our voice architects' heads, and our conversational designers' heads, turned into guides a new ADM can ramp on in a week.
• Manual config into product. Voice customization, multilingual setup, a dozen other knobs that used to eat days of engineering time now take an ADM minutes.
• Hundreds of test calls into one Duet simulations run. We worked with voice engineering to ship the latest voice simulations, which surface the issues for us. Pair it with Duet Autopilot and the fixes land without lifting a finger.

Six months later, the whole team ships better voice agents every day. An ADM launches a voice agent on the same timeline and with the same ease as any other channel, with little to no agent engineer involved. Time-to-go-live is down ~37%. Eng hours per launch have more than halved. Better yet, customers are now building the most sophisticated outbound flows on their own.

I came to voice because I love to talk. I stayed because we turned the hardest channel we build into one anyone can. We didn't just figure out how to ride the voice tsunami. We taught everyone else how to surf it too.

Cowabunga.