First Prompt Singapore · Jan 14, 2026

No More Prompts

How Voice Agents Pass the Vibe Check
场合First Prompt Singapore
日期2026.01.14
讲者姚光华 Colin
形式国际会议主题演讲(英文)
THE ARC02
WHAT THIS TALK IS ABOUT

Vibe is not a prompt problem. It is physics.

01 · SOTA Latency wins 2026 年 1 月的天花板 02 · PHYSICS Vibe is physical 三件事会毁掉对话感 03 · SURVIVAL Three laws 别听错、别失控、别让人等 04 · MEASURE VoiceAgentEval 能被感知,就能被衡量
Nobody passes the vibe check by writing a better prompt.这场演讲只讲一件事:让人觉得「对面是个活人」的,是可以被拆开、被衡量的物理量。
ACT 01
SOTA
延迟压过智能
先看一眼 2026 年 1 月的 Voice Agent SOTA 停在哪里。它没有把北极星指标定在「更聪明」,而是定在「更快」。
01 SOTA 02 PHYSICS 03 SURVIVAL 04 MEASURE
SOTA · JAN 202604
VOICE AGENTS SOTA · JAN 2026

Where latency outweighs intelligence.

TOLAN'S NORTH STAR
P90 < 1.5s整个产品的北极星指标,是一个延迟数字:90 分位的响应必须落在 1.5 秒以内。
LATENCY IS PARAMOUNT
The frontier moved
2026 年 1 月的 SOTA 里,延迟的权重压过了智能。一个更聪明但更慢的 Agent,用户不会等它把话说完。
SOURCE · openai.com/index/tolan
当行业最好的那一版把北极星定在延迟上,说明比赛的赛道已经换了。
MONEY QUOTE · 01
Latency is paramount. Where latency outweighs intelligence.
Voice Agents SOTA, Jan 2026 · Tolan's North Star: P90 < 1.5 s
ACT 02
PHYSICS
三条物理定律
毁掉对话感的从来不是「模型不够聪明」。是三件很物理的事:等待、分心、崩掉。
01 SOTA 02 PHYSICS 03 SURVIVAL 04 MEASURE
PHYSICS · LATENCY07
OUR BRAINS HAVE A STRICT SPEED LIMIT

Waiting kills the conversation.

~ 200 ms FACE-TO-FACE Instant feedback · It feels alive ~ 400 ms REMOTE CALLS We adapt We subconsciously wait ≥ 1 s HUMAN-AGENT We tolerate Not a conversation — but command & response FACE-TO-FACE / REMOTE CALL / HUMAN-AGENT
人对延迟的容忍不是无限的,它取决于对面是谁。今天绝大多数 Agent 停在第三格:能用,但那已经不是对话,是指令与应答。
MONEY QUOTE · 02
If a human pauses, we assume they are thinking. If an agent pauses, we assume it is broken.
同样的沉默,落在人身上叫思考,落在 Agent 身上叫故障。这就是延迟的不对称惩罚。
PHYSICS · ATTENTION09
THE COCKTAIL PARTY PROBLEM

Noise doesn't kill the vibe. Distraction does.

Steady Noise AC · Fans · Machinery STATUS · SOLVED Sudden Noise Dog barking · Door slamming STATUS · MANAGEABLE Background Voices TV · Cafe chatter Others talking IMPACT · THE SILENT KILLER 噪声抑制解决不了它 因为那也是「人在说话」 A MATURE AGENT MUST KNOW WHO TO LISTEN TO
前两格已经被工程解决了,真正砸掉对话的是第三格。
PHYSICS · ATTENTION10
WE CALL THIS

Selective Attention Locking.

DE-NOISE Remove what you don't want 把不想要的声音减掉。这条路解决了空调、风扇、机械这类稳态噪声,也能兜住大部分瞬态冲击。
ATTENTION LOCKING Decide who you listen to 在一屋子人声里锁定对话对象。它要的不是更强的降噪,而是选择性注意力 + 语境判断 + 音频分离。
A mature agent doesn't just de-noise. It must know who to listen to.
ACT 03
SURVIVAL
三条生死线
听清了、也够快了,还剩最后一件事:它不能在客户面前崩。稳定性在人机对话里不是运维指标,它就是体验本身。
01 SOTA 02 PHYSICS 03 SURVIVAL 04 MEASURE
SURVIVAL · ARCHITECTURE12
WHY WE CHOOSE CASCADED OVER END-TO-END

Real vibes don't crash.

CASCADED · OBSERVABLE ASR LISTEN LLM THINK TTS SPEAK OBSERVABLE MODULAR RESILIENT END-TO-END BLACK BOX One failure, total failure.
In human-agent conversation, stability IS the experience.
SURVIVAL · ARCHITECTURE13
WHAT CASCADED BUYS YOU

Three properties you cannot fake.

01 · RESILIENT
Fail Gracefully
When one part fails, the system adapts. It never crashes.
出问题时,系统不会整体失控。
02 · MODULAR
Freedom to Swap
ASR / LLM / TTS. Upgrade or replace them independently.
三段可以独立替换、独立升级。
03 · OBSERVABLE
No Black Boxes
Every ms of latency and every failure point must be traceable.
每一段延迟、每一个失败点,都要能被定位。
End-to-end is beautiful — until the first thing goes wrong.
MONEY QUOTE · 03
In human-agent conversation, stability IS the experience.
End-to-end is beautiful — until the first thing goes wrong.
SURVIVAL · THE 3 LAWS15
THE 3 LAWS OF VOICE AGENT SURVIVAL

Three laws. Miss one and there is no vibe.

LAW 01Don't Mishear.Selective Attention Locking —— 在一屋子人声里锁定你的对话对象,而不是把听到的都当输入。
LAW 02Don't Break.Observable Cascaded Architecture —— 每一段延迟和每一个失败点都可定位、可替换、可兜底。
LAW 03Don't Make People Wait.Real-time Network —— 把响应压回人脑还愿意把它当对话的那个区间。
这三条不是加分项,是准入线。先活下来,再谈体验;三条里破一条,vibe check 就过不了。
MONEY QUOTE · 04
Don't mishear. Don't break. Don't make people wait.
The 3 Laws of Voice Agent Survival · Selective Attention Locking / Observable Cascaded Architecture / Real-time Network
ACT 04
MEASURE
可衡量的物理量
活下来之后才轮到这一步:既然 vibe 是物理的,它就能被定义、被打分、被比较。
01 SOTA 02 PHYSICS 03 SURVIVAL 04 MEASURE
MEASURE · VOICEAGENTEVAL18
BEYOND SURVIVAL

If vibe is physical, it can be measured.

SURVIVAL Don't Mishear Don't Break Don't Make People Wait VoiceAgentEval THE TURING TEST FOR VIBE & ACTION VIBE ACTION OUR CLAIM We are not just building Agents. We are defining the Physics of Human- Agent Interaction.
Once you accept that vibe is physical, you realize it can be measured.
MEASURE · WHERE WE SIT19
ENABLING REAL-TIME CONVERSATION TO TRULY HAPPEN

Apps above. Models between. We are beneath.

Apps ABOVE Models BETWEEN Real-Time Conversation AGORA · REAL-TIME NETWORK BENEATH
Enabling Real-Time Conversation to truly happen.上面那两层每年都在换,最下面这一层一旦不稳,上面全白做。
TAKEAWAY20
WHAT TO TAKE HOME

Three things to take home.

01
Stop tuning prompts
通不过 vibe check 的原因,几乎都不在 prompt 里。它在延迟、注意力和稳定性上。
02
Survive first
别听错、别失控、别让人等。三条生死线过不去,再好的人设也没人听。
03
Then measure it
承认 vibe 是物理的,就能给它建一把尺子。这就是 VoiceAgentEval 在做的事。
No more prompts. Just physics you can measure.
Thank you

No More Prompts

How Voice Agents Pass the Vibe Check
场合First Prompt Singapore
日期2026.01.14
相关档案演讲档案 09
说明依原稿 8 页重构 · 标题与金句保留英文原貌