hi there. Custom-built AI agents can be used in organizations to automate all sorts of tasks with an individual context. It is also possible to build hierarchical structures with multi-agents drafted here, which can be controlled through a single personal agent, a supervisor, which a user can interact with. Like this simple multi-agent example built with an ADN, which can manage Google calendar events and notion tasks. But let me show you this in detail later. Let's take a look at user interfaces. Typically, a chat window is used to interact with these agents, but it's a quite painful experience typing on a keyboard, especially on mobile devices. Takes ages to get things done. What is the most intuitive and natural way to interact with an agent? Your voice. So, OpenAI has this amazing voice mode where you can have pretty much a natural conversation. It only has one major limitation: you cannot call external workflows with it. Custom GPTs always require a manual confirmation, which is not possible in voice mode. Other interfaces like Open UI, Libra chat, Siri, Telegram, WhatsApp, you name it, are

either not suitable for this task, are very difficult to set up, or have a bad user experience regarding voice control. So just imagine there was an app which you and even your entire organization could easily use to connect to a custom AI agent and would offer a great audio experience. Well, lucky for you, that one already exists. It's called Vagant, a lightweight app which can connect to a custom agent using a single web hook. Let me show it in action. Could you block me some focus time for tomorrow between 2:00 and 4:00 p.m.? I found a conflict in your schedule. You are busy from 2:00 p.m. to 3:00 p.m. tomorrow. Would you like to schedule the focus time from 3:00 p.m. to 4:00 p.m. instead? Okay then, let's do 3:00 to 5:00 p.m. Here is the event. Shall I create it? Yes. You're all set, and here is the new calendar entry. So, there are a few things to notice here. Obviously, the audio quality. It uses open

AI in the background. This also enables an automatic detection of more than 60 languages, then a mix of audio and visual feedback while text and voice output can defer, especially when showing drafts. It would be unnecessary to read out loud everything. The output supports markdown, so it is possible to use indentation for example to visually separate drafts and also note that it asks for confirmation before executing, but more about that later. Now, how does this work? In the settings, you can define your OpenAI API key, which is used for the voice conversion. Then, insert your webhook URL, which connects to your AI agent and handles the request. Finally, secure the connection by header auth. You can generate a password for this. I use a random UUID here. Conversations are bound to a session ID. You can reset the chat by tapping here. Now, let's take a look at the back end. This is the multi-agent example I was showing earlier. Actually, it needs to be split out into three workflows to use it, one for each agent. So, it looks like this. Let me create a new request through

the app so we have some fresh sample data. I heard about the Cal marketing strategy. Could you create an important task for it so I can read more about it later? Would you like to add a more detailed description for the task? Yes, an effective marketing strategy which is mainly built on the poll principle, making cold emails obsolete. Here is the drafted task. Shall I create it now? Let's take a look at this main agent. It starts with a web hook which is registered in the app. The header o is defined by this credential and on web call it Returns the session ID and the transcribed user input. The next node just renames the variables before they get processed by the main agent. This tool agent uses the input and also the chat history of the active session to determine which action to perform and thus which sub agent to forward the request to. All that is defined by this system prompt. The sub agents are defined as separate workflows, each of them has a description which tells the main agent when to call them for more accuracy. The original

input from the web. Hook is also passed as a separate parameter. Let's have a look at the notion task agent within this sub-workflow. Another tools agent is called. It determines how to handle the actual request and which tools to use to perform actions. Here is the belonging system prompt. It also describes the required confirmation by the user mentioned earlier and gives an example of the output format. These angle brackets enable the indentation you saw earlier in the app. This tool takes the necessary parameters which have been collected by the agent and passes them to the notion API endpoint to create a task. The chat memory is separate for every agent to keep the context clean and avoid overlapping. To achieve this, the ID of the sub-workflow is appended to the original session ID. The output of the message is then returned to the main agent, which is told to forward the output without any modifications to fully ensure that we prepend the short prompt to the output. Now back to the main agent. So the response of the sub-agent gets processed by this tools agent again.

it removes the short prompt, which was added earlier, and outputs only the desired result. And here comes the trick to skip reading out the content of the drafts out loud. The resulting message is split right at the indentation with a little bit of code. If there is no draft, both text and speech get the same content. Finally, these two values get returned to the app as a response to the webhook call. Same for the other sub-agent. It is creating Google calendar events, which I demonstrated earlier. This one contains two tools: one for checking the availability and the other to create an event. But technically, it's not much different from the other one. So, these were only two simple examples. Just imagine what is possible with this, especially when the sub-agents begin to cooperate. To get started, just visit vent. There, you can download the app, find the documentation, and even download a multi-agent example for n8n. Oh, and this project is open source. You can find it on GitHub. Let me know your thoughts in the comments, and see you in the next one. Bye for now.

Talk with your automations

Talk with your automations