Do AI agent builders actually test whether 'who the agent is' and 'how it talks' are separate design decisions — or just assume it?

This explores whether splitting agent customization into identity decisions (name, icon, who the agent appears to be) and interaction decisions (dialog, streaming, tone, how it behaves turn by turn) rests on evidence or only on convention. The short answer is that the split's best-known form is vendor doctrine, not a tested finding. Microsoft's guidelines describe a foundational tier that sets identity up front and an extended tier that shapes interaction continuously, but the guidance comes from stated design principles rather than experiments Should agent customization be split into foundational and extended tiers?. So the corpus has no direct test of this split. It does have several pieces of indirect evidence that point the same way.

The strongest comes from how people actually perceive dialogue agents. When researchers asked users to rate conversational partners, impressions broke into three separate factors: competence (about half of the variation), human-likeness, and communicative flexibility How do users mentally model dialogue agent partners?. That doesn't map neatly onto identity versus interaction. Still, it shows users judge 'what kind of thing this is' separately from 'how well it adapts while we talk,' which is roughly the line the identity/interaction split draws.

The most surprising evidence is about timing. In partner-selection games with almost 1,000 people, disclosing that a partner was an AI first made participants avoid it. Over repeated rounds, the AI's consistent and generous behavior won people over, and it ended up beating human partners Do humans learn to prefer AI partners over time?. In other words, identity shapes the first impression and interaction shapes the lasting one. That's a real reason to treat them as separate design decisions: they act at different moments, and the second can override the first.

Practitioner evidence supports the interaction side. Fin's team found that a single warm opening message raised customer satisfaction from 72.8% to 78.4% in an A/B test, and they argue conversation design deserves its own ownership Does conversation design ownership measurably improve AI agent performance?. They also recommend a related split: tone and behavior rules go in the agent's instructions, while factual content goes in a retrievable knowledge base Should factual content and behavior rules live separately?. If you want a layered view of where these decisions sit in the wider system, Dubberly's capability stack (chatbot → system prompt → RAG → supervisors) is a useful map How should agent capabilities be organized into functional layers?.

Put together, the corpus suggests the split is useful but probably incomplete. Identity sets expectations, interaction earns or loses trust over time, and the perception research hints that competence is a third dimension, and the biggest one. Neither tier on its own covers it.