Skip to main content
Sapiom logo
Sapiom

Build and run AI agents

I led product design, UX/UI and frontend engineering through Sapiom’s $35M Series A launch. The platform brings agent customization, runtimes, tool integrations and workflow automation together.

  • AI agents
  • Workflow builder
  • Developer experience
  • Design systems

Project overview

  • Overview

    Build and inspect AI workflows

    Create agents, edit workflows and inspect runs in one workspace.

  • Role

    Founding Design Engineer

    Product design, UX/UI and frontend engineering. Rebuilt the UI, design system, website and demos; led launch creative direction. Partnered with founders and engineers on strategy, AI-native prototypes and MVP features.

    Team: Founder CEO, Engineering and AI Researcher

  • Problem

    Chat and workflow fell out of sync

    The early builder could accept a chat response without updating the workflow. Long replies buried decisions, making it hard to tell what an agent would do.3

  • Solution

    Build, revise and inspect one workflow

    Keep the conversation, editable plan and run evidence together. Show each change in context, with controls to revise the plan or stop a run.

Chat, workflow and run details together

Chat and canvas share selection; Steps keeps your place as you inspect a run.

From the workspace to a nested agent

Collapse the panels to focus on the canvas. Inspect the scheduler, open its workflow, then return to the receptionist.

From sign-up to a testable agent

Two questions turn a goal into a testable template, with a live preview of its steps.

Create a workspace

Show the product working beside the sign-up form.

From a template to a running agent

People can compare steps, tools, cost and approvals before opening a template in Studio.

Browse

Compare templates by their steps, cost and whether a person signs off.

From building an agent to operating it

The web console gives teams a shared entry point for Agent studio, Router and Runtime. Agent activity, usage and connected accounts stay accessible as work moves from an individual workflow to a fleet.

Start in one workspace

Choose a product, continue an agent or start from a template.

Agent studio on mobile

Chat, workflow and run activity each have room to work on a phone.

Start from the promise

Lead with what the agent will do and show a real plan before asking for an account.

Shared components and code

I built shared React components in TypeScript and Tailwind and documented their states in Storybook, covering workflow steps, folder navigation, streaming replies, human approvals and run controls.

Visual & token foundations

Surface colors, typography and shared workspace primitives across Agent Studio.

Components in context

A local specimen. Inspect its appearance without running an action.

Approve live run

leasing · local test passed

Waiting
Tools
Your accounts
Estimated cost
$0.03

Inspect the system

Control state
Status
Surface--surface-raisedComponent surface
Action--inkPrimary control
Waiting--amber-textStatus tint

Default appearance · waiting status. Preview only.

Six roles, one palette

  • Canvas
  • Surface
  • Action
  • Complete
  • Waiting
  • Failed

Three type roles

Heading

Chat, workflow and run details together

Interface

Build with a clear plan. Review each step before your agent runs.

Telemetry

credit.check · run_0821 · 0.8s · $0.02

Spacing rhythm4px

Control heights32 / 40 / 44px

Corner radii8 / 12 / pill

Token and component referenceValues, sizing and variants

Color roles

Role and tokenLightDark
Canvas--surface-base#f4f4f2#0b0e13
Surface--surface-raised#ffffff#12161d
Action--ink#141417#f4f4f5
Complete--green-text#0f7a40#8ce3b0
Waiting--amber-text#8a5800#eec584
Failed--red-text#b02a26#f4a09d

Selected source values shown in a local component specimen.

Type and geometry

Geist and Geist Mono

Heading
16px · 500 · -0.015em
Interface
14px · 400 / 500 · 1.5
Telemetry
12px · mono · tabular
Spacing

4 · 8 · 12 · 16 · 20 · 24 · 32 · 40 · 48 px

Control heights
  • Compact · 32pxToolbars and dense desktop panes
  • Default · 40pxButtons, fields and the composer
  • Touch · 44pxEvery mobile target
Corner radii

8px · 12px · Pill

One control family

The same shape, action color and focus rule across labels and icons.

  • Workflow
  • Chat
  • Code
  • Trace
  • Approve
  • Folder
  • Copy
  • Run

Engineering practices

Small implementation choices that hold the system together.

  • Semantic colors

    Widgets read semantic tokens, so light and dark swap in one place.

    Implementation
    background: var(--surface-raised);
    border: 1px solid var(--ui-line);
  • Tints from one source

    Status fills mix from their text token and keep contrast in both themes.

    Implementation
    background: color-mix(
      in srgb, var(--green-text) 12%, var(--surface-raised)
    );
  • Widgets size to their slot

    Ledger and composer reflow by container width, inside the case study or the app frame.

    Implementation
    container-type: inline-size;
    @container (max-width: 420px) { … }
  • Visible keyboard focus

    Ink rings appear for keyboard users; pointer clicks stay quiet.

    Implementation
    button:focus-visible {
      outline: 2px solid var(--ink);
      outline-offset: -3px;
    }
  • Motion that can stand still

    Run steps and panels settle instantly when less motion is requested.

    Implementation
    @media (prefers-reduced-motion: reduce) {
      animation: none;
    }
  • Steady, announced status

    Costs and durations hold their width, and run state reaches screen readers.

    Implementation
    font-variant-numeric: tabular-nums;
    <span role="status">…</span>

Research behind the workflow

Mock-server walkthroughs and a review of eleven developer and agent tools informed planning, approvals and run inspection.
PatternReferencesDesign decision
Plan review before executionParse the request into an inspectable step plan first. Ask a question only when the answer changes a card, and keep confirmed choices editable before the build.
Synchronized chat and workflowBind the conversation and canvas to one shared ledger so every answer updates the graph in the same turn and delegates cleanly to child agents.
Step-level trace and cost inspectionExpose per-step inputs, outputs, latency and metered tool costs directly on the workflow rather than burying them in a separate log stream.
Explicit authority and approval gatesSeparate local stubbed test runs from live cloud execution, and pause at a human approval gate before any external write or paid capability call.
Patterns synthesized from product research and internal walkthroughs.3, 10, 11, 12, 13

Plan review before anything runs

The request becomes a step plan that can be checked and changed before the build. Every clause stays traceable to the step it created.
  1. One clause, one step

    Each part of the request maps to a visible step, so nothing disappears into the build.

  2. Ask only when it changes a step

    The review threshold decides routing in step 05, so it is the one question worth asking.

  3. Cost before commitment

    Step count and estimated cost per run sit beside the build action, and every choice stays editable.

Approval before a run leaves the sandbox

A live run stops at a gate that states what it will touch and spend, in plain language, next to the test it passed.
  1. Test and live never look alike

    Stubbed tools at no cost on one side, your accounts and an estimate on the other.

  2. Name what leaves the sandbox

    External writes are listed with the tool that performs them, so the risk is specific.

  3. Approve something concrete

    Scope, spend and test results are checked first, and the decision is recorded with the run.

ProductDocumented capabilityProduct implication
Gumloop5Guided setup, connector accounts, per-tool approvals and task-level credit costs.Every capability call needs explicit access scopes, action boundaries and cost attribution, not only a prompt.
Zapier6Copilot builds and revises automations in context while keeping building and testing separate from live task usage.Keep draft configuration, local stubbed testing and live production execution visually distinct.
Mastra7Workflow graphs, step execution, tool outputs, errors and repeatable evaluation datasets.Make it straightforward to inspect a failed or degraded step, revise the plan and re-run locally.

Visibility is more common than evaluation

Some observability
89%
Offline evaluations
52.4%
Online evaluations
37.3%
LangChain’s survey of 1,340 respondents; Anthropic’s evaluation guidance and Gartner’s 2025 forecast (projecting over 40% of agentic projects canceled by 2027 due to cost and risk controls) reinforce exposing what ran, what changed and what each step cost.4, 8, 9

Preparing the launch demo

I built repeatable Agent studio and Router demos for recording and hands-on exploration. The sequence followed one task from agent creation to execution and usage costs. An early 52-second concept film tested the story before final production.
Initial concept film, 52 seconds.

The concept film, in six scenes

An illustrated summary of the early 52-second film.
  1. $
    00:03The cost problem
  2. ModelsBrowsersSandboxesPayments
    00:15Remove the barriers
  3. Email verified prospects…
    00:28Describe the outcome
  4. Verified prospect outreachGeneratingSearch the domainFind the right contactsVerify every addressEnrich with company dataSend and record
    00:31Build the workflow
  5. Verified prospect outreachTimeCostSearch the domainFind the right contactsVerify every addressEnrich with company dataSend and record
    00:36Inspect execution
  6. Ship agents,not infrastructure.Start Building
    00:49Ship agents

One node, every state it can reach

Border, badge and time change together, so a run reads correctly in greyscale and at a glance.

Following one run on a phone

On mobile each plane gets the full screen. Tabs switch the view, never the context: the selected step travels with you.
  1. Chat

    Approve from the message. Open step jumps to the same step in the other views.

  2. Steps

    Scan status and cost. The waiting step carries the same Hold badge as desktop.

  3. Canvas

    Structure on demand. A half-height sheet keeps the graph visible above it.

Later demo storyboards

September 13, 2026 studies for the agent map, step inspection and run evidence.
Composition study adapted from a September demo storyboard showing a project rail beside an agent map
Plan the agent map
Composition study adapted from a September demo storyboard showing input, output and execution state in a step inspector
Inspect a step
Composition study adapted from a September demo storyboard showing a run receipt with cost and model-routing evidence
Read the run evidence

Three products, one platform story

Each product page leads with one promise, then shows the product working: a routed request, an agent taking shape in Studio, and a run that waits for approval before it resumes.
Sapiom homepage hero with the headline Ship AI agents, not infra, and the Agent Studio workspace below it mapping a research brief into steps on a canvas.
The homepage pairs one promise with the real workspace.

Router

One endpoint, with a model chosen for each request.

Agent Studio

The request, the workflow and the source behind each step.

Runtime

Run evidence and the approval that lets work continue.

Product launch and results

Sapiom introduced Router, Agent Studio and Runtime alongside its Series A. Design and development continued after launch.
Series A1
$35M
Launch-post views2
2M+
Media placements2
67
The final launch film, published by founder Ilan Zerbib.

Sources and notes

13 references
  1. Sapiom
    The next trillion agents. Sapiom’s product launch and $35M Series A announcement.
  2. Internal launch recap
    Company-reported launch results. Views are not unique people; 67 placements include 55 syndications. These are company milestones, not measured effects of the interface.
  3. Project records
    Project walkthroughs, team reviews and merged design engineering contributions. Interactive examples use sample workflow data. Read the text case.
  4. LangChain
    State of Agent Engineering, published June 12, 2026. Public survey of 1,340 respondents, November 18 to December 2, 2025. Some observability 89%; offline evaluation 52.4%; online evaluation 37.3%.
  5. Gumloop
    Agent configuration, connector accounts, tool approvals, task history and credit costs. Current product documentation, reviewed October 18, 2026.
  6. Zapier
    Copilot supports conversational automation building and revision. Building and testing do not use live plan tasks. Open beta at review, October 18, 2026.
  7. Mastra
    Studio supports workflow graphs, step execution, tool outputs, errors, datasets and experiments. Documentation reviewed October 18, 2026.
  8. Anthropic
    Demystifying evals for AI agents, January 9, 2026. Evaluating an agent includes its execution record and the resulting state of the environment.
  9. Gartner
    June 25, 2025 forecast: over 40% of agentic AI projects canceled by the end of 2027. A projection citing costs, value and risk controls, not an observed failure rate.
  10. Anthropic
    Introducing Claude Code, Anthropic. Demonstrates explicit task decomposition, tool-use and token accounting, and keyboard interruption.
  11. OpenAI
    Introducing the Agents API, OpenAI. Demonstrates streaming session events, structured evidence tables and human mitigation approval.
  12. Linear
    Linear for Agents announcement, May 2025. Demonstrates issue delegation to an agent, status transitions and inspectable work links.
  13. Google Cloud
    Design agents in Agent Studio, Gemini Enterprise Agent Platform documentation. Covers visual workflow mapping, real-time response testing and event inspection.

More work