On this page
Today uRun opens to everyone, and we’re introducing the first model on it: Sense, a decision model built for perception. Sense is trained to follow live streams and answer structured questions about what is happening in the moment.
Is the person finished? Are they asking to interrupt, or simply acknowledging what they heard? Should the agent keep listening?
You know the feeling of talking to someone who is really listening. You can pause to find a word without losing your turn. A quiet “mm-hmm” is enough to tell them to keep going. When you change your mind halfway through a sentence, a person understands and waits for you to complete the thought.
But today an agent interrupts while you are still thinking. It mistakes encouragement for an interruption. Or it leaves you waiting long enough to wonder whether it heard you. You start adjusting how you speak to accommodate the software.
We believe agents should adapt to the people talking to them.
Sense hears the audio itself rather than a transcript, so it can use tone, pace, and hesitant pauses alongside the words.
Sign up at urun.sh to try Sense with free credits.
The human timing behind conversation
Conversation happens on a short timescale. In a study of question-and-answer exchanges across ten languages, the most common turn transitions in each language clustered around 200 milliseconds, with meaningful differences between languages.
We manage this timing by preparing while we listen. Research on turn-taking describes a mismatch: 200 milliseconds for turn transitions, while producing language can take more than 600 milliseconds. People anticipate where a turn is going before it ends.
Timing also carries social meaning. A 2022 study found that people felt more connected during conversations with faster and appropriately timed responses. When researchers changed only the gaps in recorded conversations, outside listeners judged the faster versions as more connected.
For us, that makes timing a central design problem for AI. An agent can give a correct answer and still make a person feel unheard.
Many voice-agent stacks handle agent steering with silence timers, voice-activity thresholds, and hand-tuned rules. These are useful measures, but a timer alone cannot tell you why someone paused. The same interval of silence can mean “I’m done,” “I’m thinking,” or “I’m looking something up.”
There is no single millisecond threshold that makes an agent feel human. Sometimes the right response is immediate. Sometimes it is giving someone room to think. The system needs to keep interpreting what is happening so it can choose the right moment.
That is exactly what we are building Sense for: it lives in the moment.
Agent control and steering that works across modalities
Inspired by Jev and the team at Typesafe who built it, we trained a decision model for multimodal inputs. Sense returns structured scores and probabilities. You define the questions, choose the thresholds, and connect the answers to actions. Your code controls what happens next.
When an agent interrupts, you can inspect the question, score, and threshold that triggered it, then tune the behavior without replacing the model that generates its reply.
Sense works alongside your existing reasoning, retrieval, and generative models. The uRun platform supports multiple models in the same session, with low latency between them.
What it looks like in code
The SDK registers your questions and returns answers in waves: one complete answer set per tick. You can read each wave or register handlers that fire when conditions become true.
This excerpt assumes escalate is a handler defined by your application:
from sense import Sense
app = Sense(hz=10) # evaluate the question set 10 times per second
wants_refund = app.noul(
"Is the customer asking for a refund?", id="wants_refund"
)
urgency = app.score(
"How urgent is this message?",
["Low", "Medium", "High"],
id="urgency",
)
session = app.session()
session.text.push("Incoming audio is a customer support call")
# Fire at 0.8 refund probability with Medium or higher urgency; reset below 0.6.
session.when(
wants_refund.p.between(0.8, 0.6) & (urgency >= 1),
escalate,
)The .between() call adds hysteresis, so a score hovering near a threshold does not repeatedly flip your agent between states. This excerpt shows question registration and a handler; the real-time guide covers connecting the live audio stream.
The same handler pattern can start or cancel background work, such as canceling a drafted reply when the user takes the turn.
Start building with Sense
Sign up at urun.sh to try Sense with free credits. Install the Python or TypeScript SDK and follow the real-time guide to open a Session and act on decisions as they stream in. For a quick first test, the question types guide also walks through a single request. Share your referral codes to earn free credits together, or upgrade to the Team plan to invite teammates to your organization.
Sense is live in the US West and US East regions today. If you are building elsewhere, contact us about adding a region.
Sense is the first Model API on uRun. We will be adding more models on a rolling basis, and we would love to hear what you need next.
Tell us what you build. Find the SDK at github.com/urun-sh/sense-sdk.