How an AI Agent Turned a Live Stream into a Real‑Time Interactive Show for 935,000 Viewers

On June 30, a Douyin live stream titled “女流 66” showcased an AI‑driven interactive game where three participants collaborated in a physical space while the AI acted as scriptwriter, host, scene controller and task scheduler.

The broadcast integrated phone camera, microphone, room objects and audience interaction into a single game logic, turning gifts into “gold rain” and letting viewers influence subsequent game scenes.

By breaking the traditional one‑way live‑stream model, the AI enabled a new paradigm where host, AI and audience co‑create content in real time.

Running such a multi‑person, multi‑content, long‑duration interactive session revealed three core uncertainties:

Input uncertainty: multimodal signals (voice, video, gestures, gifts, comments) are often ambiguous and change rapidly.

Temporal uncertainty: actions of the host, audience and tasks are tightly time‑dependent, so a change in one second can affect the next.

Output uncertainty: generative AI outputs are probabilistic, yet the live show requires low latency, stability and safety.

To lock these uncertainties, Bagelive built an Agent Runtime that enforces high‑strength constraints, validation, graceful degradation and real‑time audit.

The runtime relies on an omni‑model (oLM) for multimodal scene understanding, S2S low‑latency voice‑to‑voice conversion with intent recognition, SSOT and ordered event broadcasting for state convergence, and SLA‑driven multi‑model orchestration to keep tail latency low.

Fault‑tolerance mechanisms include automatic fallback paths, graceful degradation of unstable model outputs, and real‑time state re‑alignment when the host deviates from expected flows, all designed to be invisible to users.

Metrics from the two‑hour debut show 935 000 total exposures, a peak of 29 000 concurrent viewers, and interaction counts 6–7 times higher than comparable streams, demonstrating both audience conversion from spectators to participants and stable, zero‑incident operation.

Beyond engineering, the team emphasizes a “program‑effect Agent” that monitors and shapes audience emotion and relationships, using emotion‑aware modeling and reward signals derived from multimodal interaction data to avoid reward hacking.

The product’s long‑term moat is described as a closed loop of Runtime + Scene Intelligence + Benchmark + Reward, which generates proprietary data, evaluation standards and feedback that continuously improve the system.

Bagelive’s AI‑native methodology restructures complexity management, development processes and organizational collaboration, treating AI as foundational infrastructure that handles context, verification and continuous optimization.

Overall, the live‑stream serves as a proof‑of‑concept for a generalizable AI Agent Runtime capable of powering digital employees, AI hosts, social AI and multiplayer games in complex real‑world environments.