Build AI Voice Agents That Don't Break in Production: Developer Workshop

Every voice AI demo works. Most production deployments don't.

A lot can go wrong on call number 3,000. A customer might phrase their query differently than your prompt expected. State scattered across databases can cause tool calls to fire out of order. The LLM might invent prices that aren't on the menu. When something fails, you have five vendors to debug and none of them show you what actually happened on the call.

This workshop teaches a different approach. Join our head of Developer Experience for a hands-on experience where you will build voice agents on an infrastructure where state lives inside the call, logic lives in your code, and the platform shows you exactly what happened on every turn.

Food and drinks are on us. You will walk away with a working agent in your GitHub.

• Downtown Chicago, IL
• Wednesday, June 24th, 2026 from 5:30 PM - 8:30 PM CST
• Every attendee is entered to win an iPad at the end of the workshop.
• Venue and parking details will go out by email and in your calendar invite.

1. Prompt stuffing. Why "you are a helpful agent that follows these 47 rules" will always lose against context drift. Learn what to put in the prompt and what to put in code.
2. No state management. Why state belongs inside the call session, not in your Redis. Learn how to make transfers carry context automatically.
3. Ignoring the latency budget. Why a few seconds of silence destroys the caller experience and how to measure the full voice pipeline and keep latency low.
4. Tool calls as an attack surface. How giving LLMs too many responsibilities creates unpredictable system behavior.
5. No post-call observability. See exactly what happened on every call and why. We'll use the open-source post-call analysis tool that ships with the SignalWire Agents SDK.

What You'll Build

By the end of the workshop you’ll have a working voice agent in your own GitHub, built on patterns you can take to production:

• An AI-driven IVR with the SignalWire Python SDK, voice and messaging on one flow
• Step-based agent logic where the LLM only sees the tools relevant to the current step
• Memory and context that persist across the call without external state management
• Function fillers so the caller hears keyboard typing while your tool calls execute, instead of dead air
• Post-call observability piped to an open-source dashboard that shows every turn, every tool call, every barge-in, and every millisecond
• Deployment patterns that scale from 10 calls to 10,000

Meet Sigmond. A live AI agent built on SignalWire. Ask him about our platform, AI architecture, or FreeSWITCH. He can search docs, answer questions, and even collect your info for a follow-up. This is real-time voice AI in action.

Agenda

• 5:30-6:00pm Check-in, networking, refreshments, environment setup
• 6:00-6:35pm Welcome and introduction
• 6:35-8:00pm Workshop
• 8:00-8:30pm Q&A, raffle, wrap-up

The Thesis

AI Agents you can count on in production require consistent behavior across tens of thousands of calls. Let the LLM handle what it’s good at: tone, intent, and data extraction. Your code should handle business rules, calculations, and what the agent is allowed to do next. AI Agent behavior should be testable, reproducible, and auditable. 

Who This Is For

Developers building voice, telephony, AI, or contact center tools who are tired of demos that work and deployments that don't. Comfortable in Python, comfortable reading API docs, ready to write code that works in production.

Why SignalWire

SignalWire was built by the team that wrote FreeSWITCH, the engine that powers Zoom Phone, Five9, Vonage, and Amazon Connect. Twenty years of carrier-grade telephony, with an AI kernel running inside the media engine instead of bolted on through webhooks. The platform owns the state of every call: who's on it, what step they're at, what the agent is allowed to do next. Your code controls the agent. Everything else is handled.