skip to content
all work
02/2023 → 2026shipped

Thread AI

Founding Product Designer

  • AI
  • Agentic
  • Workflow
  • Builder
  • Orchestration
  • Enterprise
  • 3yearsMay 2023 to Aug 2026
  • 18engineerssupported as sole designer
  • 2funding roundsshipped through both
  • 1designerthe first one hired
00

Problem

Agent systems are non-deterministic, so the work they do is hard to see and harder to trust. As the first designer, the job was to make that behavior legible to the operators running it, not just to the engineers who built it.

01

Approach

  1. 01Designed the primitives, the canvas interaction model, and the mental models that make non-deterministic agent systems legible, inspectable, and trustworthy to operators.
  2. 02Shipped production front-end code, owning components in the codebase and opening PRs directly rather than handing off specs, which removes a translation layer between design intent and what users get.
  3. 03Built and maintained the design system and token architecture as a single source of truth spanning Figma and code, and designed the brand from scratch across product, marketing site, commercial assets, and documentation.
02

Outcome

Took Lemma, Thread AI's platform, from zero to product: its surfaces, the underlying design system, and the company's visual identity, as the only designer through 18 engineers and two funding rounds.

03

The canvas

Worker Builder is where an operator composes an agent workflow and watches it run. The interaction model came first: a graph readable at a glance, the library on one side, configuration on the other, and test runs streaming underneath, so a change and its consequence stay on the same screen.

Worker Builder: composing and running an agent workflow
The same canvas in dark theme, test runs streaming below
Full width, with the expression editor and context panel open
04

Making a run legible

Agent systems are non-deterministic, so trust has to be designed rather than assumed. Configuration sits beside the node it belongs to instead of behind a modal, and every run leaves an inspectable trace. The wireframe underneath is how much structure that legibility actually costs.

Agent configuration beside the node it belongs to: model, credentials, prompts
Wireframe x-ray of the builder, every panel and rail in one view
05

The primitives

Under the canvas sits a small vocabulary: call, switch, parallel, handoff, wait, loop. Each primitive carries its own inline controls, so whatever you want to do to a node is on the node. Get that vocabulary right and complex orchestration composes out of parts an operator already understands.

Primitives: plan generation, handoff, wait, and a tool loop
Dynamic islands, the contextual toolbar for each node type
State primitives: every node type with its inline controls
06

The system underneath

Every surface above draws on one library, maintained as a single source of truth spanning Figma and code. I owned those components in the codebase and opened PRs directly rather than handing off specs, which takes out the translation layer between design intent and what users actually get.

Agent and group components, expanded and collapsed
Navigation: collapsed rail, expanded menu, console settings
Panel system: library, preview, call, and handoff configuration
Component library: worker flows and agent panels drawn out in full
Worker Builder 2.0 component sheet, primitives through panels
07

Brand, out of the same tokens

Tokens are where the design system and the brand meet. One palette, defined once and themed across light and dark, then carried out of the product into the marketing site, commercial assets, and documentation.

Color tokens across light and dark
08

The exploration behind it

The prototypes these surfaces came out of are still drivable: the node canvas, the function registry, worker states, the API editor, and the rest of the run. They are exploration rather than shipped product.

playground

Thread AI Playground

The design exploration behind the AI workflow builder — the node canvas, function registry, worker states, and more.

Available for select product design work.

If the problem in one of these looks like yours, I would like to hear about it.

LinkedIn