What design rules improve AI agent usability in practice?

Matthew Phillips, writing on Salesforce's blog, reports that Salesforce's internal product and UX teams codified lessons from building the company's own Agentforce platform — used internally as "customer zero" before being sold to customers — into an internal framework called the Agentic Interface Guidelines. The post lists five "actionable insights" from that framework: limit agent names to under 10 characters using a "Name + Agent" convention; use three heuristics — consistency, error tolerance, and teachability — to catch usability problems early; give every agent conversation a five-part structure (a welcome message, a disclosure statement, a list of example questions, clear scope boundaries, and a path to give feedback or escalate); always include AI disclosure language; and standardize visual branding, including voice and tone and agent avatars, so that "a consistent look signals to users: 'You're talking to an AI agent and here's what to expect.'"

The framework draws on existing resources rather than inventing fresh guidance — it borrows from Salesforce Lightning Design System 2's Agentic Experience Patterns, Salesforce's own branding and marketing style guides, and materials from Slack — and each tip is tied to a concrete failure the team hit while building Agentforce. Long or vague agent names, the post says, "broke layouts, cluttered dropdowns, and confused users," which produced the naming convention. Agents that "looked good on paper but confused real users" prompted the heuristic checks. And "a blank input field is intimidating" motivated the welcome-message-plus-example-questions structure. Phillips frames the disclosure requirement as a trust move rather than a compliance one: "AI isn't perfect, and users know it. But we still need to say it."

Set against How should AI systems be designed across different interaction phases?, Salesforce's framework covers similar ground — onboarding expectations, error recovery, consistency — but is narrower and more concrete, and it is validated only by Salesforce's own internal dogfooding rather than by the practitioner study behind Amershi's guidelines. It also overlaps with What UX principles do workplace users want in AI agents?: Salesforce's scope-boundary and disclosure tips read as a concrete implementation of that framework's context-awareness and safety principles for workplace-facing agents. And it offers a thinner version of the expectation-setting half of How can users verify what an AI agent actually plans to do? — a welcome message, scope boundaries, and example questions tell a user what an agent is supposed to do, but the framework as described stops at setting expectations up front and does not address verifying what the agent actually did.

The post is vendor content from Salesforce's own blog about its own Agentforce product, written by a member of its design team who writes "we" throughout; it reports no user study, no metrics, and no outside validation of the five tips beyond the anecdote that a reimagined Salesforce Help agent experience "launched it in five days" after one observed failure. The 10-character naming limit and the three named heuristics are stated as internal design conventions, not tested thresholds, so the specific figures should be read as Salesforce's own design-system defaults rather than general findings about agent usability.