Can frontends handle delegation while staying conversationally engaged?
How can a conversational interface continue interacting smoothly while delegating long-running tasks to background systems? The challenge is interpreting results in a conversation that may have evolved during execution.
Realtime-Venus presents delegation as part of the interaction design of a full-duplex system, not as a tool call that pauses it. Each of its two 9B models, Realtime-Venus-Omni for audio-visual interaction and Realtime-Venus-Audio for spoken interaction, is "a complete conversational frontend" that integrates continuous perception, conversational control, and native speech generation. A "dual-loop runtime" then lets foreground interaction continue while Realtime-Venus-Harness "executes tasks asynchronously and returns results for integration into the ongoing dialogue." The discussion concludes that coordinating immediate conversational responses with longer-running computation is "a promising direction" for assistants that stay engaged while handling tasks beyond the frontend's own capabilities.
The introduction gives the reasoning. Continuous interaction and external computation "operate on different timescales within the same session." A background task needs a stable record of the request and its supporting evidence, yet its result must be interpreted in a conversation that may have changed while it ran. The paper's answer is a shared causal timeline that holds user inputs, model outputs, and delegation events together. The delegated request and the later result then sit in the same ordered record as everything said in between, so the frontend can place the result in the conversation as it now stands.
This extends Can a single model learn when to speak and respond?. That note argues that timing and turn control become learnable once perception and generation share one causal stream. Realtime-Venus reads as adding delegation events to that stream, so the unification covers work handed off, not only speech and video. It also stops short of a single model: the background reasoning and tool execution live in a separate harness, and the frontend is explicitly not expected to do that work itself. It is also a concrete interaction-level version of the persistence argument in What makes an AI system feel like a colleague rather than a chatbot?, since the record of the request and evidence carries a task across a changing conversation. The paper's framing of proactivity, responding "without an explicit user request", bears on Why can't conversational AI agents take the initiative?.
The excerpt does not test the delegation design directly. The numbers it reports are video benchmarks, where Realtime-Venus-Omni scores highest among the evaluated online models on six of eight, including StreamingBench (70.2%), OVO-Bench (64.7%), and Daily-Omni (81.3%), plus improvements from memory augmentation on hour-scale video across all evaluated duration bins. Claims of "strong conversational continuity under non-interruptive speech" and "competitive multimodal understanding" come without figures here. It reports no delegation success rates, result-integration quality, or latency, and it does not name the comparison models or say how the two frontends were trained beyond a "common post-training recipe". The paper itself lists timing as open, with finer-grained streaming chunks as future work. What the evidence supports is the architecture as a coherent proposal backed by strong frontend perception results, not a measured advantage of asynchronous delegation over blocking tool calls.
Inquiring lines that read this note 21
This note is a source for these research framings, grouped by the broader line of inquiry each explores. Scan the bold lines of inquiry; follow any specific question forward.
How can conversational AI develop appropriate initiative and goal awareness?- How should a frontend interpret delegated results after the dialogue has shifted?
- Can a single streaming model handle turn-taking better than cascaded pipelines?
- How does wait time and interruption frequency interact with backchannel effectiveness?
- Do task-specific interfaces outperform conversational chat in practical settings?
- Do specialized interfaces outperform generic chatbots for domain-specific work?
- Why do dynamic UIs reduce cognitive load but complicate user control and predictability?
- What UI patterns beyond chat complement AI for task-oriented workflows?
- Should hybrid interfaces default to clicking or conversation for structured tasks?
- Can UI and conversational chat coexist without making users feel interrupted?
- Why do some studies claim UI outperforms chat while others treat them as concurrent?
- How do task-oriented UI patterns compare to friction within single chat turns?
Related concepts in this collection 3
This note in its neighbourhood — explore the map, then jump to a related concept in the list below.
Click a node to walk · click center to open · click Open in graph to see this note in the full knowledge graph
-
Can a single model learn when to speak and respond?
Does combining perception, generation, and turn-taking into one streaming model let timing and interruption handling emerge naturally, rather than requiring separate engineered modules?
extends: adds delegation events to the shared causal stream while keeping background work in a separate harness
-
What makes an AI system feel like a colleague rather than a chatbot?
This research explores whether colleague-like AI requires bigger models or better architecture. It investigates which design features—persistence, memory, reusable skills, task closure—actually drive the shift from episodic tool use to sustained work partnership.
convergent-with: a persistent record of request and evidence lets a task outlive a changing conversation
-
Why can't conversational AI agents take the initiative?
Explores whether current LLMs lack the structural ability to lead conversations, set goals, or anticipate user needs—and what architectural changes might enable proactive dialogue.
qualifies: the paper targets responses initiated without an explicit request but reports no proactivity-specific results here
Related papers in this collection 8
Papers most semantically related to this note, ranked by cosine similarity in the embedding space.
- Realtime-Venus: A full-duplex interaction system with asynchronous delegation
- Delegating or Doing? Understanding User Behavior in Hybrid Human-Agent Interfaces
- LLMs Corrupt Your Documents When You Delegate
- Disclose with Care: Designing Privacy Controls in Interview Chatbots
- Don't Break the Cache: An Evaluation of Prompt Caching for Long-Horizon Agentic Tasks
- When Single-User-Oriented LLM-based Assistants Involve Others: A Scoping Review of Pathways, Risks, and Responses
- When Patients Cut In: Extending Clinical Conversational AI Safety to Interruptions
- Probing How Users Interact with Turn-Level Design Frictions for AI Chatbots
Original note title
a full-duplex frontend can delegate asynchronously and keep interacting — Realtime-Venus shares one causal timeline across inputs, outputs, and delegation