Role:
Primary designer
Team
UX researcher, PM, Dev
Timeline:
2025 - 2026
to create an API with Agent, down from ~3 days manually
monthly queries, +23% month-over-month
of sessions ended in a successful outcome
CONTEXT
The problem wasn't capability. It was complexity.
Developers could create, test, troubleshoot, and manage APIs in API Connect BUT these workflows involved many steps, configurations, and tools strung together. Early research showed that manually creating an API could take roughly two weeks. AI presented an opportunity to rethink how developers moved through this complexity—not simply add another chatbot.
How might we design an AI Agent that turns complex API workflows into simpler, intent-driven interactions while keeping developers informed and in control?
RESEARCH INSIGHT
Discover
Create
Secure
Test
Manage
Socialize
Developer, usability session
Developer, usability session
DEVELOPERS WANTED
BUT WORRIED ABOUT
HOW I WORKED WITH RESEARCH
DESIGN DECISIONS
01.
Too much confirmation created friction. Too little undermined trust. So I split Agent actions into two types: read actions, like search and inspect, run immediately since they carry no risk. Write actions, like create and deploy, show a plan first, so developers always know what's about to happen before it does.
02.
Once the Agent started working, silence created uncertainty. I designed a lightweight progress trace to show what’s happening, what’s done, and whether it’s still working without exposing every backend step.
03.
I separated explanation from action: chat explains, the UI applies.
When a developer asks the Agent to fix an error, the change is applied directly in the editor and highlighted with a brief gradient that fades after it’s been seen. This gives developers clear proof of what changed without adding permanent clutter.
The pattern became the foundation for subsequent Agent experiences.
08 — FROM FEATURE TO SYSTEM
The flagship wasn’t the feature. The interaction model was.
API creation was the flagship workflow, but the larger opportunity was a framework that could support many kinds of AI-assisted work. The same model extended across discovery, security, testing, and troubleshooting — different workflows, same mental model.
DISCOVERY
Prevent API sprawl before it starts
Search existing APIs and reusable assets before creating something new.
SECURITY
From reactive fixes to proactive guidance
Review APIs against organizational patterns and security practices before deployment.
TESTING
Generate coverage from API intent
Generate test cases, execute them, surface failures, and recommend fixes.
TROUBLESHOOTING
Investigate in chat. Resolve in context.
Diagnose issues conversationally, then return the actual fix to the API workspace.
09 — SCALING ACROSS PRODUCTS
One interaction model across the ecosystem
The first Agent experience launched in VS Code. I extended the interaction framework into API Connect and later API Studio. Each surface had different workflows and constraints, but the model remained consistent: start with intent, understand the plan, follow the work, act on the result in context.
API Manager
45%
API Studio
30%
VS Code
25%
What began as a solution for API creation became a reusable product pattern.
10 — OUTCOMES
Adoption showed where the model worked — and where it didn’t
78% of sessions ended in a successful outcome; 8% required escalation; 14% were abandoned. The abandonment rate mattered — rather than hiding it, I treated it as a roadmap signal for where longer or more complex workflows still created uncertainty.
TOP CAPABILITY
API Discovery
28% of all usage — reinforcing the research finding that developers wanted help understanding existing systems before creating something new.
COMMON WORKFLOW
Discover → Doc → Code
32% of multi-tool sessions followed this exact path — developers using the Agent to understand before they acted.
RETENTION
4.2×
Average repeat use within 30 days.
11 — WHAT I LEARNED
Trust is an interaction system
No single piece creates trust alone
A plan alone doesn’t create trust. A trace alone doesn’t create trust. A result alone doesn’t create trust. Together — predict before acting, see during execution, review before accepting — they do.
Patterns create more leverage than features
The broader impact came from establishing a framework that could scale across discovery, testing, security, troubleshooting, and future Agent experiences.
Constraints can improve the outcome
The limits of chat forced a clearer separation between explanation and action. That constraint produced a better model: the Agent explains the work, the product lets developers work with it.
The gap is the roadmap
The abandoned sessions became the clearest signal for what to design next.