# Real-Time AI Agents with Jev: Fast Decisions | QQuantum.ai > Real-time AI agents that read the state and choose the next move. See what TypeSafe’s Jev Doom demo shows, how parallel questions work and how to scope a pilot. Source: https://qquantum.ai/technology/jev/real-time-agents [Home](https://qquantum.ai) [technology](https://qquantum.ai/technology) [jev](https://qquantum.ai/technology/jev) real-time-agents Jev use case · Real-time agents # Read the state. Make the move. Real-time AI agents have to decide within an interaction budget, and waiting for a long generated answer can break the experience. For product and game teams, Jev opens a different design: compact text state, several focused questions asked in parallel, and a bounded next action your code applies. TypeSafe reports 70–500 ms per request; your full loop adds network and application time. Explore a responsive agent Structured text state · Proposed applications, not a live game All Jev: AI Tools & Fast Decisions → Introduction transcript. Qubi: When a product needs a decision now, a long written answer arrives too late. We send a compact snapshot and ask several independent questions together. The answers choose the next move; if a reply is late, a safe default holds. We time the whole loop under real load before calling anything real-time. ## The demonstration ## Yes, TypeSafe showed Jev playing Doom. TypeSafe’s launch demonstration used game state represented as text and structured data. The team described making ten queries per second. This shows a model participating in a responsive game loop; it does not establish direct visual gameplay, competitive performance or a guaranteed frame rate. The provider also reports a 70–500 ms response range for System One requests, with evaluations generally run from the US West Coast. Network distance, input size and application work affect the complete loop. [Read TypeSafe’s launch and demo qualifications](https://typesafe.ai/blog/introducing-system-one-models-and-jev). For a product team, the useful question is which decisions can fit into the interaction budget. A game’s strategic action selector can update less often than its physics or rendering. A live interface can classify intent while ordinary code updates the display. ## Parallel questions in practice Same questions, fewer round trips Seconds · lower is better · Linear scale from 0 to 3.00 - **One batched Jev request**: 0.27 - **13 sequential Jev requests**: 2.71 13 questions about one document, averaged over five runs. The baseline sums sequential calls; concurrent calls would narrow the latency difference. Source: [TypeSafe parallel-questions cookbook](https://docs.typesafe.ai/cookbooks/parallel_questions). Checked 18 September 2026. Provider-reported, not independently measured by QQuantum.ai. Shared context reduces input cost USD per briefing · lower is better · Linear scale from 0 to $0.007000 - **One batched Jev request**: $0.000497 - **13 separate Jev requests**: $0.006090 The same document is sent once in the batch and repeated in separate requests. This is a batching comparison within Jev, not a comparison with another model. Source: [TypeSafe parallel-questions cookbook](https://docs.typesafe.ai/cookbooks/parallel_questions). Checked 18 September 2026. Provider-reported, not independently measured by QQuantum.ai. ## What parallel means ## Ask independent questions together. A simulation can ask which permitted move fits the current goal, whether an obstacle blocks the route, and whether the state needs a fallback. These questions can use the same snapshot. Your code combines the outputs and discards checks that are irrelevant to the selected path. If a question needs the result of an earlier action, it belongs in a later request. Parallel evaluation does not make sequential dependencies disappear. TypeSafe documents this distinction in its [speculative fan-out pattern](https://docs.typesafe.ai/patterns/fan-out). Keep the snapshot small and relevant. Persist memory in the application, pass the current facts explicitly, and reject a decision if its state version is already stale. These are proposed engineering controls for a responsive integration. ## Applications to explore PLAY ### Game and simulation agents Select from permitted moves using text state supplied by the game engine. Keep physics, collision rules and time-critical safety checks in deterministic code. INTERACT ### Responsive product interfaces Classify a request, select a relevant help path or decide whether to ask a clarifying question. Measure the full interaction time from the user’s region. SELECT ### Agent skill selection Choose an appropriate capability from an approved catalogue. TypeSafe’s skill-suggestion example ranks candidates, reads a shortlist in detail and can reject all options. Application ideas, not shipped QQuantum.ai case studies. Source for skill selection: [TypeSafe cookbook](https://docs.typesafe.ai/cookbooks/skill_suggestion). ## The response budget ## Measure the complete loop. Instrument state collection, network transit, model response, validation and action execution separately. Track median and tail latency under realistic concurrency. Test timeouts, rate limits and stale responses before describing an experience as real-time. An agent should have a safe default when a response arrives late: keep the previous permitted action, pause, or choose a deterministic fallback appropriate to the application. A low-latency average is not a hard real-time guarantee. For camera or microphone input, another component must first produce text or structured observations. The [current Jev model reference](https://docs.typesafe.ai/models) lists text-only inputs. Include that conversion in the latency, quality and cost budget. Questions ## Questions about Jev Real-Time Agents. ### Can Jev directly watch a game screen? The current documented model accepts text, not images or video. TypeSafe’s Doom demonstration used structured game state. A visual pipeline would need a separate perception component. ### Is the batching chart a game benchmark? No. It reproduces TypeSafe’s document-question experiment. It illustrates batching overhead, not gaming performance or a guaranteed response time. ### Can this replace a hard real-time controller? A hosted model call does not provide a hard timing guarantee. Keep strict timing and safety constraints in deterministic systems and evaluate whether a model belongs in a slower decision layer. Keep reading - [Model Build with Jev](https://qquantum.ai/technology/jev) - [Use case Agent Guardrails](https://qquantum.ai/technology/jev/agent-guardrails) - [Evidence Published Benchmarks](https://qquantum.ai/technology/jev/benchmarks)