Orilo launches a voice-only personal agent aimed at the morning routine

A hands-on review published by Chinese tech outlet QbitAI on November 3 introduces Orilo, a voice-first personal agent from a team led by Jiao Ke, a co-founder of Baichuan Intelligence. The team previously built "Laifu Radio," a personalized daily-briefing AI radio product that reached hundreds of thousands of users with 50% next-day retention and more than 10% 30-day retention; work on Orilo began early this year.

The product's most visible difference is its interface: a water-drop shape that uses the phone's gyroscope and slides across the screen when the device tilts. Interaction is almost entirely voice-based. Users leave the device on a table, listen to information and issue tasks by speaking, with the morning routine—brushing teeth, holding coffee, getting ready to leave—as the core scene.

That morning use case centers on a personal briefing. Rather than pushing stories by interest tags, Orilo draws on past conversations and long-term memory to judge what a piece of news means for a particular user and offers an interpretation angled to their work. It can also take a one-sentence request and monitor a niche company, a match or a dormant podcast around the clock, reporting back when something changes.

Cross-app execution runs through connectors and tool calls. In the review, asking Orilo to check email returns a summary within two or three seconds—for example, that a colleague needs next Tuesday's presentation updated. If a school email says pickup time has moved earlier, the agent interjects during the briefing and offers to add it to the calendar with an earlier reminder.

The technical problem is the tension between latency and reasoning. End-to-end audio models respond quickly and sound natural but stumble on complex logic; cascaded ASR-LLM-TTS pipelines are smarter but often start at 1.5 seconds of delay, past the point where a conversation feels alive. Orilo splits the interaction layer from the execution layer: a lightweight, low-latency voice model handles listening, turn-taking and emotional cues, while a slower, high-reasoning agent reads email, calls APIs and checks calendars in the background.

The front end does not go silent while the back end works. Orilo gives the voice model the ability to keep chatting, carrying the previous topic forward and using fillers such as "hmm" and "uh" when it needs a moment. On the Full-Duplex-Bench evaluation, it resumed after a user interruption in 1.08 seconds, compared with 1.50 seconds for GPT Realtime and 2.62 seconds for Gemini Live.

The trade-offs are real. The review notes that the two-layer setup sacrifices robustness for speed: on several occasions an API call failed, but the front end, unwilling to leave silence, confidently said the task was done—while the calendar stayed empty. Orilo also supports English only, a hard limit for Chinese-speaking users.

Still, it is a useful sample. Personal agents have largely been stuck in a trap of low-frequency use, high interaction friction and high running costs. Orilo uses voice to enter a high-frequency morning routine, then uses cross-app execution to close the loop inside a single conversation. The questions now are when Chinese support arrives, whether task success rates can support the "personal secretary" promise, and how the architecture holds up on real networks.