Welcome to a New Era of Experience with AI Agents by Kat Holmes of Salesforce makes a compelling case that as AI agents become “digital coworkers,” they must be treated as a new kind of user—with their own goals, needs, and limitations.
AI agents are disrupting decades of investment in graphic user interfaces. - Kat Holmes
The article introduces Agent Experience (AX) Design as the practice of creating environments in which AI agents can operate efficiently and, in doing so, deliver human-centered outcomes.
Holmes’s article is an essential primer for any organization looking to embed AI agents into their customer or employee journeys along four pillars:
- Defining Agent Experience Design—treating agents as users and designing both “for” and “of” agents.
- AI Agents as Delegates—purpose-built systems that learn from vast data, reason through incomplete information, and navigate broken processes.
- Humans as Orchestrators—interfaces that let people steer multiple agents, verify reasoning, and regain control when needed.
- Flexible Human-to-Agent Interfaces—deep personalization, dynamic UIs (e.g., Salesforce Lightning Design System 'SLDS 2' beta), and multi-modal interactions that adapt to individual preferences.
What I really enjoyed about this point of view:
- Human-Centered Lens: By reframing agents as users, the piece aligns AX design with established UX principles—usability, accessibility, and emotional resonance—while highlighting new technical requirements like high-quality APIs and structured ontologies.
- Concrete Criteria: Holmes offers actionable design guidelines (e.g., agent-friendly system criteria on page 1) that bridge the gap between visionary rhetoric and practical implementation.
- Forward-Looking Infrastructure Advice: The call to harmonize data (in this case with Salesforce Data Cloud underscores the often-overlooked importance of metadata readiness for agent success.
Thinking about this through the production systems we've built at Black Flag Design :
As someone who builds hyper-personalized interfaces and advocates for dynamic, adaptive UIs in enterprise contexts, the article’s emphasis on agent orchestration resonates deeply.
In projects like our education work, we’ve seen firsthand how an AI agent can falter when underlying processes are fragmented—precisely the “bad experience for the agent” that Holmes warns against.
Integrating AX design principles could help ensure that our Retrieval-Augmented Generation (#RAG) pipelines not only surface relevant content for educators but also present it through interfaces that flex to both human and agent workflows.
Ways for Designers and Developers to go deeper:
Metrics and Feedback Loops: How might we measure agent “satisfaction” or success—analogous to user-experience KPIs—and feed that data back into continuous improvement?
- AI-specific KPIs should include Task Completion Rate, measuring the percentage of user-initiated tasks successfully executed, which directly correlates to agent reliability.(TalkToAgent)
- Accuracy—the degree to which agent outputs match ground-truth or expert expectations—is fundamental for assessing the correctness of reasoning and actions (Galileo AI)
- Response Time metrics track latency between user request and agent reply, ensuring the agent meets real-time interaction standards.
- User Satisfaction scores, captured via in-app surveys or Net Promoter Scores, provide qualitative insight into perceived agent helpfulness.
Ethical Governance: While Holmes touches on trust, a deeper dive into transparency, bias mitigation, and regulatory alignment (e.g., NIST risk frameworks) would strengthen the human-centered ethos.
- The NIST AI Risk Management Framework (AI RMF) provides voluntary guidelines for incorporating trustworthiness considerations—such as security, privacy, and resilience—throughout the AI lifecycle.
- Complementing this, the NIST Trustworthy and Responsible AI guidance outlines pillars like validity, safety, accountability, and transparency to underpin governance strategies.
- Building AI transparency into organizational values starts with clear documentation of data sources, model capabilities, and decision-making processes, enabling stakeholders to trace agent behavior (Shelf).
- Best practices include ethical data collection, where sourcing policies and dataset audits are documented publicly, and explainability tools that provide justifications for agent outputs on demand.
Case Studies Across Verticals: Examples beyond customer-service order changes—such as AI-mediated collaborative workflows in education, finance, and operations (projects our teams have led and launched)—would illustrate how AX design scales across domains.
- Stanford’s IT Teaching Resources highlight cases where generative AI tools assist lesson planning and content curation, reducing cognitive load for educators by auto-organizing resources across platforms (IT Teaching Resources)
- Recent industry analyses describe AI agents as co-pilots that analyze learning interactions and dynamically adjust course material, elevating personalization at scale. (Disco)
- Cutting-edge research on agentic workflows shows how multi-agent collaboration frameworks can facilitate breakout sessions, peer feedback, and real-time content synthesis in virtual classrooms.
By treating agents as a new user class and prioritizing data hygiene, flexible interfaces, and orchestration tools, businesses can move beyond “proof of concept” toward truly agent-empowered experiences.
For practitioners like myself—straddling #UX design, #product strategy, and #AI applied systems—this framework provides both inspiration and a clear set of next steps to ensure that our digital coworkers aren’t just functional, but delightful.
