One agent per topic, written in your language.
It is not a single chatbot with a tree of options. It is several virtual people, each with a role, and the model picks which one takes the message.
Conext agents are not a menu. You write the tone and what they may promise. The CRM picks who speaks, and you can change it.
Name, role, instructions. No code.
Each agent has a name you see in the inbox, a role (who they are for the business), and the instructions you write: tone, what they may promise, what they may not. You turn them on and off. You reorder them. You delete them, always after a confirmation.
The auto-send switch belongs to each agent, not to the whole CRM. Sales can send list prices on their own; complaints leave the text for a person to look at. That ceiling always wins: if it is off, the model does not go out to WhatsApp even if it feels sure.
The model picks who speaks. You can change it.
On every inbound message, Gemini picks the agent and writes the reply in the same step. If the contact already had one, that is context; if the model names one that does not exist, the one already on the thread stays.
You can change the agent by hand in the thread. The panel shows if it is off. Opening an agent from the bar opens its settings, not another filter: it is a different screen, with how many conversations it is handling.
In practice
- 1
You create it
Name, role, and instructions. Auto-send off at first if you do not know it yet.
- 2
It handles the chat
The model picks it when the topic fits, or you assign it from the thread.
- 3
You tune it
If it gets ahead of itself, you drop it to draft. If it already replies well, you let it send.
Agents do not replace the team. They take the work you already know how to answer.
Want to see it with your own threads?
Essential includes one channel, 500 AI conversations per month, and one active agent.