Breaking

Arena engineers race Agent Mode from napkin sketch to taco-truck waitlist app

September 5, 2026 at 11:59 EDT

AI Agents
Software Dev & Coding

Two Arena engineers, Graham and Logan, each turned the same napkin sketch into a working taco-truck waitlist app in Agent Mode, with a build-versus-buy call on the line. The company is inviting others to run the same prompt against leading models on Agent Arena.

September 2026 · Arena

From napkin sketch to waitlist app

Two Arena engineers, Graham and Logan, each built the same taco-truck waitlist app in Agent Mode from one hand-drawn sketch, with a build-versus-buy recommendation on the line.

How the brief moves

1 · SketchOne napkin drawing

→

2 · Agent ModePlan, tools, sandbox

→

3 · App + callWaitlist plus build vs buy

What people already ask agents to do

Coding is the largest reported mix in Agent Mode; each column is one percentage point per block.

29%

Coding

11%

Research

11%

Planning

~4%

Workflow auto

Blocks are proportional (about 4 points each). Coding towers over adjacent categories.

What landed

A working waitlist app from one shared sketch, plus a chance to rerun the prompt on Agent Arena.

What still binds

Users often keep a tight leash; the orchestrator stays unnamed; GitHub push stops after a merged or closed PR.

Leaderboard backdrop

2,188,416 Agent Arena sessions across 16 labs as of September 1, 2026.

The napkin race sits on a live evaluation of how orchestrators handle coding-heavy, tool-using work—not a one-off bake-off with a published winner.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.