Voiceflow V4: The New Context Engine

Key Takeaways

Latency is the number one killer of AI agent adoption. You can build the smartest agent in the world, but if it takes 5 seconds to respond, customers won't use it. They'll escalate to a human, they'll send you emails, and they'll churn. And in the enterprise, this whole problem just gets harder. Your agents are connected to CRMs, ticketing systems, and payment platforms. Real, often complex infrastructure that doesn't always respond quickly, especially at scale. So, we rebuilt VoiceOps' already leading agent execution engine to be even faster. We call it the context engine. The context engine is the runtime behind every agent conversation. It streams responses token by token in real time. So, customers hear or see answers as they're generated, allowing for blazing fast, human-like conversations. And it's natively multimodal, a single API that handles both voice and chat. What makes it a context engine is that it's constantly synthesizing everything it needs to respond. Your business data and tools, the customer's context, the conversation history, its short and long-term memory, and the situation at hand. All in real time, all while delivering responses simultaneously. The agent doesn't stop to think, it reasons, retrieves, and responds all at once. But sometimes, speed isn't up to your agent's context engine alone. In enterprise environments, your agent needs to work across complex systems, and not every API responds in milliseconds. So, we reimagined how agents work with tools. Tool calls can now run asynchronously. The agent keeps the conversation going while work happens in the background. No dead air, no awkward pauses, and when a tool call is unpredictable in length, you can make custom messages that fire on start, on completion, or on delay, so your customer always knows what's happening. The result, no matter how complex your systems, the conversation always feels seamless. That's the context engine. Now, let's talk about the models that you can run on it. We don't believe in one model fits all, and at enterprise scale, that belief becomes simple math. When an agent is handling millions of conversations, every optimization matters. You don't need PhD-level intelligence to say, "Hi, how can I help you today?" But that's exactly what happens when you run a single large model across every task. You're paying for reasoning you don't need, and you're adding latency that your customers can feel. In V4, you mix and match your models across your agent. Use a lightning-fast model like Anthropic's Haiku 4.5 for greetings, routing, and everyday tasks, and then bring in OpenAI's ChatGPT 5.2 with reasoning only where you need it, for complex playbooks that demand deeper intelligence. The right model for the right job every step of the conversation. At scale, that optimization compounds. Every penny saved per interaction becomes meaningful savings across millions of conversations. But, if you're not sure where to start, every new project comes with Voiceflow recommended model stacks, curated for your use case, balanced for cost and performance, and always kept up-to-date. You can also hover over any model to see exactly what it's best suited for. That's V4's model library. Build best-of-breed agents that are faster, cheaper, and smarter where it counts. But, here's what most people don't account for when forecasting model costs for agents. The cost doesn't scale linearly, it compounds. You don't just pay for each message, you pay for every message that came before it. Every turn, every tool call, every handoff adds to your agent's context window, and that window gets carried forward with every response. The longer a conversation runs, the more expensive and slower it gets. That's why memory management is such a critical part of V4's context engine. First, the framework itself is memory optimized. When your agent moves from one task to the next, the context engine hot swaps instructions, loading only what's needed and clearing what isn't. No supervisor, no middleman, just a clean, lean context window at every step. Second, your agent never dumps a full tool response into the context window. It selects only the fields it needs, so a thousand-line API payload might just be a few words in the conversation. That API call that triples your context window, it doesn't happen on Voiceflow. Third, you control how much conversation history your agent holds in short-term memory. Slide it up for complex, multi-step conversations. Dial it back for high-volume, fast-resolution use cases. You decide where the trade-off lands. Then, there's long-term memory. Your context engine extracts key details from short-term memory and stores them separately, so your agent stays intelligent even after the conversation history fades. Conversations aren't limited by your context window. They can go on as long as you need, and you control the lifecycle. Persistent forever, wipe after every session, or define custom reset behavior. The result? An agent that stays fast and cost-efficient as conversations get longer and more complex. Not one that gets slower with every turn. That's the context engine, the models, the tools, the memory. Now, let's hear all of it in action. Hi, welcome to Shift. Going by your phone number, am I speaking with Tyler? Yep. Great, how can I help you today? Uh I'd like to make a payment, please. One sec. So, I can process that payment, I just need to verify a couple of things. Could you punch in the last four of your social then pound? That's 1 2 3 4, and can you confirm your date of birth? Yeah, it's uh February 18th, 1995. Okay, February 18th, 1995. Let me look you up. Great, Tyler. I've got you verified. Pulling up your account now. Now, let me grab your loans. I've got a loan here with Greenwood Automotive for $1,200. Your next payment of $110 is due on February 16th. But, how much would you like to pay? Um let's just do 110 then. Perfect. Let me pull up your payment methods. I've got a Visa ending in 9876. Want to use that? Yeah, that'd be good. Thank you. Just so you know, credit cards have a 2.99% processing fee. Still want to use that? Uh yeah. Can you tell me how much that'll be on top of my payment? Sure. The fee on $110 is $3.29. So, your total will be $113.29. Okay, cool. Yeah, let's go ahead with it. Before I process this, I need to read a quick disclosure. On Monday, February 9th, 2026, you, Tyler Han, authorize Shift Technologies to charge your credit card ending in 9876 for a one-time payment of $113.29 on Monday, February 9th, 2026. Please confirm by saying yes. Yes. Done. $113.29 has been processed, and you'll get a confirmation email shortly. Is there anything else I can help you with today? Nope, I think I'm good for us today. Thank you. Thank you for calling Shift. Goodbye. That's a real voice agent. Built on V4, and deploying one has never been easier. Embed a voice or chat agent directly on your website or mobile app, and now provision a phone number directly from inside Voiceflow. Or, bring your own telephony with custom SIP trunking. Web, mobile, phone. Live in minutes. That's V4 under the hood.