For the first year Donna shipped, it was a single agent with a bunch of tools and a long prompt. The team gave the model access to the CRM, the calendar, the email, and everything else a field sales rep might need, and then hoped the model was smart enough to pick the right tool at the right time. For most queries it was. For a meaningful slice of them, it was not. Simple user requests were triggering ten or twenty tool calls. Hallucinations crept in. Conversations that should have taken fifteen seconds stretched to a minute. The team knew the architecture was not going to scale.
The second half of Gheerwijn Clicque's talk at Gathering #13 was about what they did about it. Donna now runs on two agents, not one. A lighter conversation agent handles the spoken interaction in real time. A heavier execution agent does the deep work after the call ends. The decision sounds architectural. It is also a product decision, and it changed how the team thinks about where intelligence belongs in a voice-first AI product.
