Work/Arcade Control Plane
Case study ( 04 ) · Arcade.dev
Turning Arcade from an infrastructure inventory into an agent deployment control plane.
A take-home redesign of Arcade.dev. I chose the problem, set the north star, scoped the build and shipped a working prototype of the path to a first authorized tool call.
Context
Arcade wins when the first tool call works.
Arcade is the auth and tool-calling layer for AI agents. The take-home was open-ended, so I treated it like week one as a founding designer: no PM, no spec. I owned the problem, the scope and how to measure it.
Before designing anything, I tried every product flow myself, logged where it broke, and shadowed an AI engineer for 30 minutes.
North star: 1
First successful, authorized, reviewable tool call.
How I worked
Product flows tried All (Logged wins and failures); AI engineer interview 30 (Minutes, with shadowing)
Shipped: A deck and a working prototype in code.
Figma and shadcn/ui first, then Cursor with the Figma MCP.
Problem
Users meet the architecture before the value.
The dashboard is a resource inventory. Users have to assemble Arcade's runtime in their heads before anything runs.
Old dashboard: ≈8k
Tools exposed by default.
Sidebar: 18
Flat items, named after backend objects.
First decision: 3
Paths to pick before any value: code, gateway, sandbox.
Diagnosis
01 Resource inventory (18 items); 02 Decision paralysis (3 paths); 03 Fragmented setup (4 places); 04 Weak feedback (No next step)
Strategy
Build one path, completely.
Everything served one moment. I built the path to it end to end and labelled the rest.
Sidebar built: 4/12
Plus onboarding and the code wizard, outside the sidebar.
Built
Dashboard (4 states); Playground (Chat + Execute); Tool catalog (39 integrations); Servers (Gateway wizard)
Deferred: Users, policies, runs, audit and usage wait.
They matter, but none gets a new user closer to the first call.
Onboarding
Intent before infrastructure.
Ask what they want done, recommend the fastest path, keep expert paths one click away.
Before
1 Land on a resource dashboard; 2 Pick code, gateway or sandbox; 3 Assemble docs, config, OAuth; Hope the first call works
After
1 Pick an outcome (5 goals); 2 Choose a path (1 recommended); 3 Connect & run (Inline auth); Trace, then Project Home
Three paths
▶ Playground, ⇄ MCP gateway, </> SDK, ↗ Explore first
Expert wizards: 8 steps
Gateway and SDK each, with live status beside every step.
Default: Sandbox first.
“Sandbox runs are safe to test. You can promote to production later.”
Playground
A runtime simulator, not a chat box.
Every run shows the same eight checks. When a tool needs access, the run pauses and asks in place.
Checks per run: 8
Paused at 3
The trace: Paused for one scope: issues:create.
User intent; Tool selected; Authorization (Now); Scope check; Policy check; Tool execution; Result returned; Audit log saved
To authorize: 1 click
“Continue to Linear”, inside the run.
In the Playground
Modes 2 (Chat · Execute); Tools ready 64 (Across 20+ apps)
Project Home
One page, four states.
The dashboard follows the user's progress, and every blocker ships with its fix.
States
0/3 Empty, 1/3 Needs auth, 3/4 First run, 4/4 Active
First run: 3/4
Setup done. Next: enable test users.
Success rate: 91%
Blockers, ranked
Salesforce auth expired (Fix now); Slack missing chat:write (Reauthorize); Slack approval pending (Review); Gateway has no client (Copy config)
Pick tools (2 mins)
Choose the tools your agent can call.
Connect auth (3 mins)
Authorize once, inside the run.
Test action (1 min)
Run your first reviewable tool call.
Measurement
Time to a safe agent in production.
Scale it if time to first call drops without new auth drop-off. Rethink it if users run but never open a trace.
Prototype target: a new user reaches a first authorized, reviewable tool call in a single session, without leaving the product for docs.
Activation funnel
Start 100, M2 84, Auth 68, M1 54, M3 42, M4 30, M5 20. Illustrative shape. No data until it's instrumented.
Five metrics: Measured by safe production use, not time in the dashboard.
M1 Time to first authorized call (Activation); M2 Setup completion by path (Onboarding); M3 First trace opened (Understanding); M4 Repeat call in week 1 (Retention); M5 Invited user activation (Rollout)
2 weeks · Validate
Built: Prototype, tested with 5–10 agent builders.
4 weeks · Beta
Next: Onboarding and Playground MVP, instrumented.
8 weeks · Production
Later: Live node canvas, real-time runs, readiness score.
Outcome
A working path to the first call.
Simulated data, no production results yet. The metrics above decide whether it scales.
The call I'd want to own as a first designer: pick the moment that matters, build the path to it, and be honest about what waits.
Risk: The guided path hides power.
Mitigation: “Explore dashboard first” stays one click away.
Risk: Sandbox success doesn't transfer.
Mitigation: First run ends in “Enable test users”.
Risk: Simulated data.
Mitigation: the 2-week plan tests with real agent builders.