This case study is password-protected

Enter the password to view this project.

← Back to home

Internal tool redesign · Tempus Gen AI platform

Agent Builder: from Swiss Army knife to guided experience

Agent Builder is Tempus's internal, no-code tool for building AI agents. It was powerful enough that some agents outgrew their creator and spread across teams — but so open-ended that most new users didn't know where to start. I led research and design on turning it into a guided flow that surfaces only what a new agent needs first, without losing the depth advanced users relied on.

Company
TempusInternal Gen AI agent platform
Role
UX DesignerResearch → concept → wireframes → hi-fi design
Research
10 power users1-hr interviews, week of Apr 28, 2025
Testing
8 internal usersLow-fi wireframe & guerrilla usability testing
Agent Builder · Case study

Summary

A powerful internal tool that only power users could drive

10power users interviewed
8internal users in usability testing
5/10agents seen already spreading beyond their creator

Agent Builder is an internal tool that empowers teams across Tempus to create AI agents for research, operations, and internal automation. Over time it gained a reputation as a "Swiss Army knife" — incredibly powerful, but difficult to use. People could configure genuinely complex logic and behavior, but without any guidance, most struggled just to get started.

The goal: transform Agent Builder into a guided, approachable experience that helped new users build agents with confidence, while keeping the depth that advanced users relied on.

01 — The problem

A lot of power, not enough guidance

Agent Builder exposed every advanced option at once, with nothing to tell a new user where to start.

01

Where new users got stuck

  • A lot of advanced features, not enough guidance
  • Users didn't know where to start
  • A steep learning curve for anyone new to the tool
The original Agent Builder create-agent screen: a 90-day activation warning banner across the top, and a red retriever error for a missing required field, both visible in the basic editor
The default Create Agent screen. A 90-day activation warning and a technical validation error ("missing value for required field collection_id") are both surfaced in the same basic editor a first-time user lands on — exposed complexity with no explanation of what it means or why it matters.

02 — The research

Ten power users, one shared story

We interviewed ten power users across Tempus who frequently interact with AI agents. By examining their workflows in Agent Builder and other Gen AI tools, we set out to uncover the biggest friction points, adoption barriers, and opportunities for improvement — insights that shaped how "Meaningful Adoption" of Gen AI was defined and prioritized across the organization.

Research goals

  • Inform prioritization — which Agent Builder features would drive "Meaningful Adoption" of Gen AI
  • Define personal value — how Agent Builder empowers people to build "Personally Useful" agents
  • Map current tool usage — understand participants' Gen AI tool habits, to guide documentation on tool selection

Defining "Meaningful Adoption"

Before the first interview, the team defined Meaningful Adoption of Gen AI as movement toward two things: a measurable efficiency and quality impact on people's actual work, and proactive, user-driven agent building rather than passive tool use. That definition is what the research was measured against.

Methodology

Date — Interviews were conducted the week of April 28th, 2025.

Format — Qualitative research via one-hour, structured interviews conducted over video conferencing with recordings. Interviewees used screen sharing to explain the composition of their agents.

Approach

  • Understand the user's general use of various internal and external Gen AI tools and their primary applications
  • Discuss a specific Agent Builder creation, exploring its function, value, and potential improvements
  • Ask about the user's outlook on Gen AI at Tempus and gather feedback on enhancing Agent Builder's capabilities and adoption

Note — participant names throughout this research are fictional, and quotes were lightly edited for brevity and clarity.

"Streamline the Agent Builder UI for easy and guided Agent Instructions, adding data, and LLM customization."

A recommendation synthesized from the interviews

Value Derived from "Personally Useful" and "Desired Wide Usage" Agents

Half of the agents we saw in the user interviews ended up being more than just "Personally Useful": five were rated Personally Useful, and five showed Desired Wide Usage — already being developed or refined with the goal of expanding beyond their creator to a wider team.

Agents were classified by two signals: how many questions they'd been asked, and how many unique users relied on them. Agents with heavy, repeated use from a single person were labeled "Personally Useful" — and that cohort became the research focus, since their higher engagement made them the clearest signal of what makes Agent Builder valuable.

Automated Processes

Streamlining repetitive or complex tasks, reducing manual effort and minimizing errors.

Time Savings

Handling routine or time-consuming work. This efficiency frees up valuable time for other priorities.

Enhanced Quality

Improving the accuracy, consistency, and quality of outputs by applying best practices and reducing human error.

Gen AI Education

Providing a space to tinker and experiment with general Gen AI functionality as well as Agent Builder specific functionality for learning purposes.

Mapping the agent lifecycle from prototype through to monitoring surfaced where in that journey users hit the most friction.

Agent development lifecycle

PrototypeOne Internal BuildAgent Builder TestAir - Monitor EducationDocs & Courses GovernanceAgent Builder SupportCo-Pilot ShareServiceNow PublishAir - Register DeployAir - Execute IntegrateOne in TCL MonitorAir - Monitor

Where the friction concentrated

BuildAgent Builder TestAir - Monitor EducationDocs & Courses

Owners of "Personally Useful" agents experienced barriers during these initial phases — Build, Test, and Education. Those barriers prevented participants from progressing to later phases in the process, like Share and Publish, and from achieving wider agent adoption.

03 — The process

Concept, wireframes, then high fidelity

From the research, three phases: a fast concept prototype to align leadership, low-fidelity wireframes tested with real users, then high-fidelity screens refined against what that testing surfaced.

Concept development

Created a vibe-coded prototype to explore what a more guided and approachable experience could feel like. I presented it to design directors, PMs, and engineering leads — and received strong support and alignment around the vision.

The vibe-coded prototype. A fast, AI-generated concept — agent configuration on the left, a live chat preview on the right — used to align design directors, PMs, and engineering leads before any real design work began.

Wireframes and testing

Once leadership aligned on the concept, I moved into low-fidelity wireframes that mapped out the guided flow step by step. These wireframes focused on helping users understand where they were in the process and what each decision meant for their agent.

Low-fidelity wireframe of the Agent Home screen: a table of existing agents with name, description, and owner columns, an All/Mine tab toggle, a New Agent button, and a yellow usability-testing scenario note
Agent Home. A simple list of agents and one clear way to start a new one.
Low-fidelity wireframe of the Configure Agent tab: name, description, and agent instructions fields, an optional knowledge section, a collapsed Advanced configuration row, and a live agent preview panel alongside
The Configure tab. Name, description, and instructions, with a live preview alongside as you type.

Wireframes — advanced configuration, registering, and sharing

We tested these early screens with 8 internal users — a mix of people who had built agents and people who hadn't. The results were encouraging: users found the new design intuitive and far less intimidating than before, making it easy for them to get started. However, several participants struggled toward the latter half of the building process when it came to understanding what it meant to register an agent, promote it to production, and share it.

Heatmap table scoring how well 8 usability-testing participants understood each part of the wireframed flow, from configuring an agent through registering and publishing it, alongside insights and design opportunities for each row
What the wireframes tested well — and what they didn't. Configuring and adding knowledge landed clearly; registering, versioning, and publishing did not — the exact confusion the next round of design had to resolve.

High-fidelity designs and guerrilla testing

With the core flow validated, I moved into high-fidelity mocks to refine the visuals and interactions. Each step was made clear and lightweight to reduce friction and guide progress. During earlier testing, users showed confusion around the final steps — registering, sharing, and promoting agents to production — so I made focused design tweaks to clarify these actions, including a guided entry point for sharing.

04 — What changed

From open sandbox to guided flow

Surface only the core building blocks

The final designs surfaced only the core components of an agent up front, introduced inline help, and built a guided experience for sharing and registering an agent. Together, these changes turned Agent Builder from an open-ended sandbox into a structured, easy-to-use building experience.

Before

Every advanced option exposed at once. No starting point, no guidance, a steep learning curve for anyone new to the tool.

After

Only the core building blocks surfaced up front. Inline help throughout. A guided flow for sharing and registering an agent.

05 — The final design

The shipped experience

Screens from the live product — the guided build flow, and the guided flow for sharing and registering an agent that replaced the confusion guerrilla testing surfaced.

Guided build — home, configure, knowledge, LLM

Guided entry point — sharing & registration

06 — The impact

From intimidating to usable

The redesigned Agent Builder had an immediate and visible effect across teams. What had once been seen as powerful but intimidating became a tool that people could actually use confidently.

Early testing after rollout showed that new users could create a functional agent in minutes, compared to hours with the previous version. Teams reported that they needed far less guidance from technical peers, freeing up time for engineering and research staff. Stakeholders and leadership praised the new experience for striking the right balance between flexibility and simplicity — giving users power without overwhelming them.

07 — What's next

Where the work continues

  • Add the ability to equip agents with tools, so they can use them when and how they see fit
  • Research and understand why users switch from the basic editor to the advanced workflow editor
  • Surface an easier way for users to add API calls to their agents
  • Continue iterating based on user feedback and maintain a user-centered design philosophy
Next project Data Producer Studio →