Can familiar software terms gain new meaning for AI agents?

Salesforce Design's glossary post argues that AI's new vocabulary is mostly old vocabulary wearing new meaning: "many of the most important concepts aren't actually new. They're familiar words that have taken on new meaning as intelligence becomes embedded in the products we design." It defines ten terms — Agentic, Agent behavior design, Multi-agent orchestration, Context, Prompt, Memory, Ontology, Headless, Guardrails, and Explainability — grouping them under four questions designers must now answer: "How should AI behave? How does it understand? How does it earn trust? And how do we build with it?"

Each entry traces the same shift from execution to agency. "Agentic" marks the move from software that "waits for instructions" to systems that "reason, act, and pursue goals... with less direct supervision." That shift turns former UX and engineering terms into design material: "prompt" becomes "a design surface that influences how AI behaves, responds, and collaborates" rather than a one-time instruction; "guardrails" are "the constraints that define the boundaries of AI behavior," setting what an agent can access and when it must escalate to a human; "headless" no longer names an architecture pattern but the decoupling of "intelligence from the interface so AI can operate consistently across channels," to the point that "the conversation itself becomes the interface." Memory and ontology split what a system retains (context that persists across interactions, which "transforms interactions into relationships that evolve over time") from what it understands (the relationships connecting concepts, objects, and data).

The Context entry — AI systems "infer context from conversations, behavior, history, location, and surrounding data" rather than requiring users to supply it through navigation or forms — restates, in practitioner vocabulary, the same shift How does AI context differ from conventional software context? already names, but nests it inside a wider relabeling that also covers prompt, memory, ontology, and headless. Its framing of "agent behavior design" and guardrails as explicit design responsibilities — deciding when an agent asks for clarification, escalates, or needs approval before sending a customer email — runs parallel to How should AI systems be designed across different interaction phases? and to Does human-in-the-loop design actually enable effective oversight?, though it specifies no way to verify an agent is honoring the guardrails it is given. The orchestration entry's chain of "a sales agent," "a scheduling agent," and "a compliance agent" handing off work echoes the layered surfaces How should AI interfaces handle the shift from doing to supervising? formalizes.

The glossary offers no evidence that these redefinitions are shared industry-wide or stable over time — it is one design organization's working vocabulary, illustrated with its own product examples (a CRM account, a support ticket, a sales email) rather than tested cases. It poses its four questions — behavior, understanding, trust, building — without resolving any of them, and says nothing about how a designer should decide where a guardrail belongs. The glossary is best read as a snapshot of shifting practitioner language, not a validated method for designing agent behavior.