How does artificial intelligence really become smart?

They make telephone life easier for thousands of users by automating call follow-up, if not completely taking over. But there's more to it! Of course, for our customers, the focus is on which features we can provide for them and what added value they have for their daily work. But especially when (further) developing AI products, the “how” is just as important as the “what.” Because this has a strong impact on the quality of the results.

AI Assist is moving to the Agent Network

So far, each of our AI Assist features stands alone. Not only in terms of content, after all, sipgate app users are presented with the results in various places, but also technically. Based on the transcripts, every other functionality is executed, regardless of whether further AI-generated content is created. They don't “know” anything about each other. And why should they? In traditional software development, of course, individual features also have an indirect influence on each other, but AI features or products open up completely new opportunities here. Instead of putting all features side by side independently, we have now connected them together in an agent network.

An agent network (or agent-based architecture) essentially consists of several specialized “agents,” i.e. software components or AI modules, each specialized in specific sub-tasks. These agents collaborate, share information, and complement each other, rather than a single monolithic component trying to cover all features alone.

You can think of it as a team:

• Who transcribes an agent
• Another one recognizes topics
• Another creates to-dos and prepares subject lines,
• Yet another person generates conversation suggestions or answers, etc.

Better results through greater flexibility

Okay, but that sounds like exactly what we already have in terms of features, doesn't it? Yes, and this is exactly where we come back to the meaning of “how”: Agent networks offer unique advantages:

1. Better focus through context isolation

Each agent on a network has exactly one task. This must be described very clearly and differentiated from others. This allows you to formulate the instructions for the LLM in a targeted and concise manner, which has a positive effect on the quality of results. The more complex a prompt is and the more implicit and explicit instructions and subtasks it contains, the greater the risk of ambiguous results and hallucinations. Agent networks make it easy to break down tasks in a coherent structure.

2. Better quality through feedback

Because of their structure, agent networks are very flexible. The individual agents are in contact with each other, can give each other feedback and identify and correct each other's mistakes. This influence can be strengthened and deliberately manipulated to achieve even better results. In this way, you can also control the autonomy of the respective networks and individual agents.

3. Better fit through faster development

In addition, individual agents can be added more quickly and easily, which perform very specific functions, such as dedicated monitoring of results. In the past, we've had problems with summaries in incorrect languages or occasional hallucinations. Control agents can specifically identify such errors and make corrections without us having to implement complex logic. In the same way, we can also implement new functionalities in a significantly shorter period of time, without having to (build) a new API route in each case.

4. Better features through testing options

Developing AI-based features and products that really generate added value requires faster feedback loops and more detailed testing, even at early stages of development. Agent networks make this possible. Agents can easily be duplicated and adapted so that the better option can then be selected and further developed based on A/B testing. We can also implement specialized feedback agents to analyze the results and make suggestions for improvements.

5. Better results through context & tools

Last but not least: Agents are much easier to equip with tools. Be it to use information from external sources to be able to recognize and correct names and email addresses directly, or even to push specific information to the right places, such as CRMs or your own calendar.

The Agent Network becomes a powerful assistant for AI Assist

When a call ends today, it is no longer just a single function that starts in the background, but an entire team of specialized agents. First, the conversation is transcribed, which is then checked and refined by other agents. On request, a data protection agent removes personal information such as telephone numbers or addresses, while a quality agent checks and improves the transcript for typos and incomprehensible passages. At the same time, a specialist word agent ensures that industry-specific terms, foreign words and proper names are correctly entered. Even unclear dates such as “next Tuesday” are translated into concrete data by another agent so that there are no misunderstandings later on.

On this improved basis, the actual assistance functions then come into action: summaries are created, to-dos are recognized, subject lines suggested and topics are marked. Another agent then checks whether information twice or tasks appear multiple times and ensures that the result remains lean and consistent. Finally, external tools come into play: CRM systems can be automatically filled with data, calendars can be populated with appointments, or ticket systems can be provided with new tasks — all orchestrated by agents in the background.

We are currently working on the finer details to make the agent network in the labs area available to experimenting users of the sipgate app.

‍