I think starting with one agent makes more sense, especially when the product is still growing. It gives you room to understand what users actually need before splitting the system into specialized agents.

The interesting part for me is the idea of managing the agent through natural language. If it can actually understand the agent’s context and make configuration easier, that could remove a lot of the friction involved in maintaining AI systems.

The knowledge-retention point is interesting too. If an AI agent can keep the right context and processes available even when team members change, that could solve a problem that automation alone doesn’t address.