The infrastructure you do not see on the screen
During the visit to Agora.io in Shanghai, the conversation covered two problems that often seem separate: transporting audio and video over the internet, and building a useful interaction with artificial intelligence. The encounter showed how these layers come together when someone needs to converse with a voice agent.
A well-produced voice attracts attention in the first demonstration. In actual use, other questions start to matter: does the agent listen to the end? Does it notice a correction? Can it stop speaking? Does it use the right business information? Does it know when to hand the situation over to someone? The experience depends on these decisions as well as the quality of the model.
The presentation connected network engineering, conversational behavior and business operations. For product builders, the central lesson is to examine the entire interaction. The response is not complete when the model generates a sentence. It needs to arrive at the right moment and help the person accomplish something.
A company specializing in connecting people
Agora, Inc. brings together two businesses with distinct focuses: Agora serves the United States and international markets; Shengwang is headquartered in Shanghai and focuses on China. They are operations of the same holding company, and this organization helps explain the group’s presence in both markets. Products, contracts and local conditions need to be considered in their own context. The company’s 2025 annual report.
This specialization is connected to the career of Tony Zhao, also formally identified as Bin Zhao, one of the founders. Before Agora, he worked on audio and video at WebEx, founded NeoTasks and was CTO of YY. His history is tied to interactive communication challenges, rather than solely to the recent adoption of AI models. The founder’s institutional biography.
The most recent public results provide a reference for scale: in the second quarter of 2026, the group reported consolidated revenue of US$40.4 million, year-over-year growth of 18% and 3,892 active customers. The customer metric counts accounts with more than US$100 in revenue over the preceding 12 months and excludes Easemob. These are group figures, not figures for the office visited alone, released on August 13 in unaudited results. Official second-quarter 2026 results.
The shortest route is not always the fastest
One analogy presented to the group was road traffic: choosing a route depends on conditions along the way, not just distance. On the internet, packet loss, congestion and fluctuating connections also change the experience. In a conversation, a delay can cause two people to start speaking at once or break the continuity of a response.
Agora’s SDRTN is described as a real-time transport layer that monitors network conditions, selects routes and uses redundancy. Its role is to support interactive audio and video exchange. This requires attention to problems that a video watched with a few seconds of buffering can conceal more easily. How Agora describes its real-time network.
The perceived speed of an agent, however, does not depend on this network alone. The device, local connection, speech recognition, model, speech synthesis and systems consulted all contribute to the process. Measuring just one stage does not reveal how long a person waits for a useful response.
Speaking is one capability. Conversation requires others
During the encounter, the discussion of voice moved into attention, intention and turn-taking. A pause can mean that someone has finished or is still thinking. An interruption can be an important correction. The interaction needs to accommodate these situations rather than behave like a rigid sequence of questions and answers.
The Conversational AI Engine connects existing models to voice interfaces. In a chained architecture, speech is recognized, the model generates a response and synthesis turns it into audio. There are also configurations using native audio models. The official FAQ clarifies that the engine does not create or train the connected model. Its purpose is to integrate and manage the interaction. The Conversational AI Engine offering.
The documentation covers both interruptions caused by user speech and interruptions triggered by an application command. Detection of the start and end of speech is also configurable. These capabilities help shape when to listen, respond and stop, but they do not mean perfect interpretation of every human intention. Documentation on interruptions and turn-taking.
One test example would be a person correcting a request while the agent is responding. The goal is to observe whether the correction enters the flow, whether the previous response stops and whether the conversation resumes at the right point. This assesses the experience, rather than only testing whether the voice sounds natural.
A demonstration does not replace a product in operation
The difference between demonstrating and operating was one of the visit’s most useful discussions. A prototype may work with a familiar script. A product needs to handle incomplete questions, outdated information, integration errors and situations outside that script. The work involves business objectives, available knowledge and continuous improvement.
One way to connect an agent to a company’s information is retrieval-augmented generation, known as RAG. The service retrieves relevant passages from a knowledge base and supplies them as context for the model’s response. This lookup is different from retraining the model. Agora’s documentation supports connecting a custom model service and describes this use of retrieved context. Model and knowledge-base integration.
To apply this, it is useful to define where each piece of information comes from, who updates it and which actions the agent can perform. A convincing voice does not automatically make an answer correct. Assessment needs to consider the content, the action performed and the outcome for the person.
| What appears in the demonstration | What needs to be observed in actual use |
|---|---|
| The voice sounds natural. | The agent listens, accepts corrections and respects turns. |
| The response arrives quickly. | The wait is acceptable under real network and usage conditions. |
| The agent knows a script. | It uses up-to-date information and handles exceptions. |
| The integration performs an action. | The action is confirmed, recorded and handled when it fails. |
| The conversation ends. | The person resolved their need or was directed appropriately. |
Cost is also a product decision
Brazil came up in the conversation as a market where economic viability needs to accompany technical capability. The thesis presented was to combine models and technologies at a cost appropriate to the use case. Choosing the most sophisticated configuration is not enough if its price makes the service a company wants to offer unviable.
For a founder, this suggests comparing cost per task resolved as well as cost per minute. A cheap conversation that does not resolve a request can create repetition, transfers and rework. A more expensive configuration also needs to justify the difference through results that matter for that case.
Testing should consider language, accents, noise, duration, usage volume and connected systems. The right option for a short demonstration may not be the right option for many interactions over a month. These are decisions to measure in the context of the product, without assuming a universal price or cost saving.
People remain part of the system
The group’s questions brought up screening and handoff to human support as part of the design. The discussion did not support eliminating people altogether. An agent can help organize a request and prepare the next stage, while a professional takes over a situation requiring judgment or a different decision.
The handoff needs to preserve context: what was requested, what has already been tried and why someone should take over. Starting from scratch after a long conversation can cancel out some of automation’s benefits. This is a product and operational requirement. Interrupting the agent’s speech does not, by itself, implement a complete handoff to a person.
The practical question is where AI helps most and where it should stop. That boundary may change as the operation matures, but it needs to be defined from the first pilot. Responsibility for the service remains even when the interface becomes an AI voice.
The interface moves beyond the screen and meets developers
The conversation also covered devices and noisy environments, expanding the perspective beyond on-screen applications. When the interface is a voice, microphones, acoustics, connectivity and the physical context are part of the experience. The same agent may require different choices in a quiet room and a busy workplace.
There is a public example of this connection between infrastructure and practical building. In March 2026, Agora and Seeed Studio announced a collaboration on AI embedded in devices, accompanied by a hackathon that planned to include Reachy Mini robots, NVIDIA Jetson computing, Hugging Face models and mentoring. This was a program announced at that time, not an outcome of this visit. The public announcement of the Seeed Studio collaboration.
For developer relations, documentation, examples and technical support can reduce the distance between understanding a technology and being able to use it. The takeaway for community builders is to create opportunities for experimentation with clear objectives. A prototype starts the conversation; support, learning and improvement help sustain adoption.
What founders can take into their next pilot
- Start with the work to be done. Define which problem the conversation should solve before choosing the voice and model.
- Test listening. Include pauses, corrections, interruptions and incomplete questions in the assessment.
- Measure the entire experience. Observe the wait for a useful response, task completion and rework.
- Maintain the knowledge. Keep content current, define the limits of actions and verify what was executed.
- Plan for people in the flow. Define when to hand over and how to preserve context.
- Assess the economics in actual use. Compare cost and outcomes against the expected duration and volume.
The Agora visit showed that the opportunity in conversational AI involves more than adding voice to a model. It takes work on infrastructure, interaction design and business operations. Combining these layers is what can turn an interesting demonstration into a useful experience for the person on the other side.