Case study · AI platform · Omnewave
AgentStudio
An agent builder for enterprise teams. Build an agent, put agents together into workflows, and ship a new version without breaking the one that’s live.

TL;DR
- 7objects in one model, from agent to run
- 4places to look, instead of four lifecycle stages
- 2ways to deploy, chosen at release
- 3AI-made sitemaps I tested mine against
DesignedThe information architecture, the versioning model and 50+ screens, from overview to deploy.
ProblemThe first version split the product into Design, Configure and Operate. People had to guess which stage held the thing they wanted, and one workflow was managed from a list row, a menu and a side panel.
What I didMade the workflow the one place to build, test, version and deploy. Reusable parts moved to a Library. Versions got rules, so a change never surprises an app that’s already calling the agent.
Context
What Omnewave is
Omnewave is WaveMaker’s agent engineering brand. It’s for enterprise teams that want AI agents running in production, with the same care as any other software. AgentStudio is where they build agents, put them together into workflows, and run them in sandbox and production.
Omnewave builds spec-first. The spec is the contract, and the code is derived from it. That shaped the product as much as the process: every workflow is also a readable spec, and my screens reached engineering the same way.
Research
Four builders, one missing half
I went through LangGraph, Crew AI, AutoGen and Muse screen by screen. Each is good at making something: LangGraph at the visual graph, Crew AI at starting from a prompt, AutoGen at teams of agents.
They’re built for the moment you make a flow. Enterprise teams live in the months after: which version is live, where it runs, who calls it, and what it costs. That half became the brief.
Problem 01
Four stages, and no obvious place to look
The first version’s top bar read Design, Configure and Operate. That’s how a team thinks about a lifecycle, not how people look for things. To change one workflow, you could start in any of the three.
How I solved it: I named the areas after things: Workflows, Library, Analytics and Settings. Everything about one workflow lives on one page. Agents, knowledge and tools are reusable parts, so they live in the Library. Members, models and keys sit in Settings.
- Design
- Configure
- Operate
- Admin
Where did the system put it?
- WorkflowsBuild, test, version and deploy, on one page
- LibraryAgents, knowledge, tools and templates
- AnalyticsRuns, quality and cost across the workspace
- SettingsMembers, models, channels and keys
Go where the thing is.

How I got there: I wrote the sitemap first. Then I gave Claude, Lovable and Figma Make the same brief and asked each for its own sitemap. I put all four side by side. Two ideas held up and went in: one Library for every reusable part, and keeping runs, deployments, audit and analytics apart, because each answers a different question.
Problem 02
Agent or workflow? Draft or version?
Before any screen, the words had to mean one thing each. “Agent versus workflow” kept coming up, and “edit” could mean changing something that was live.
How I solved it: One object model, seven words. Each screen uses them the same way.
- AgentA reusable capability
- WorkflowAgents and steps, put together
- DraftWork you can still edit
- VersionA release that never changes
- EnvironmentWhere it runs: sandbox or production
- DeploymentOne version, running in one environment
- RunOne execution
Two rules came with it. Editing a deployed version creates a draft, so nothing live changes while you work. And a version never has more than one draft, so there’s never a question of which edit is the real one.
Problem 03
Open on what needs you
A workspace can run dozens of workflows. The first question each morning is what broke, not what exists.
How I solved it: The overview leads with what needs attention, then runs, success rate, and spend against the monthly budget. Below that, each workflow is one row, with its versions folded under it, its environment and its endpoint. A warning sits on the row that needs you.

Problem 04
Cost shows up on the invoice, too late
Agents spend tokens at every step. In most builders, you learn what a workflow costs after it has run for a month.
How I solved it: Every step on the canvas shows the tokens it used, so cost is visible while you build. Start from a blank canvas with Wave AI, or drag in nodes. The side panel only ever holds the selected node; everything operational moved to the page.


Problem 05
Shipping a change could break someone else’s app
Other apps call a deployed workflow by its endpoint. Change it in place, and every caller gets the change at once, whether they’re ready or not.
How I solved it: Test on the same canvas, with chat or JSON input, so there’s no separate test page. Then deploying asks one question: update the same endpoint, or release a new one? An update replaces the live version and keeps the old config as a version. A new endpoint leaves the old version and its callers untouched.


Problem 06
“What did I change?”
A draft can sit for days. By the time it ships, nobody remembers everything that changed in it.
How I solved it: Compare the draft with the live version before shipping. Changed steps are marked Added or Modified on the canvas, and Wave AI lists every change in plain words, with the old value and the new one.

Library
Templates that say what they need
A template you can’t finish setting up is worse than none. So every template shows what it needs before you use it: the connections, the inputs and outputs, and which connection is still missing.

The spec underneath
Every workflow is also a spec
Flip any workflow from Visual to JSON. Select a step on the canvas and the JSON highlights it. Reviewers read exactly what the agent runs.

The product is built the same way. Omnewave uses spec-driven development with Spec Kit, and my screens reach the build through figma-starter, the open-source Spec Kit extension whose Figma conventions I set. Give it one Figma section, and it writes the screens, the flow map, the build order and one spec per screen. A second command binds each screen to its APIs. The tokens those specs use come from the design system I lead, in the Design system case study.
Launch
Vantage: telling the story
Vantage is Omnewave’s AI testing tool. It explores a web app like a real user and writes the tests. I designed its product UI, and I wrote the story and the script for its launch film. I also helped shape the concept, the workflow and the storyboard, and showed the team how to make the film with AI, using Claude, and ElevenLabs for the voice.
The film opens on a countdown: an hour until the release goes out, and nobody is sure checkout still works. It walks through Vantage in six steps, then ends on the same countdown a week later. You won’t be hoping.

Next
What’s next in the designs
- Before you deploy over an endpoint, show the apps that called it in the last day.
- Save a test run as a test case, and let a workspace require tests to pass before a deploy.
- Separate places for runs, release history, audit and analytics, because each answers a different question.

