Skip to main content

Initial public release · July 25, 2026

Product Bench

A matched three-arm comparison on four greenfield thin vertical slices, scored with Harbor RewardKit.

Lamina mean
0.689
vs direct
+0.104
vs plan-first
+0.130
Tasks led
4/4

Results

Product behavior scores and task-level medians from three full reruns per task and arm.

Mean scores

Direct agent
0.586
Lamina agent
0.689
Plan-first agent
0.559

By task

Borrowing things among neighbors

dev-loan-library

Lamina led
Direct agent
0.621
Lamina agent
0.750
Plan-first agent
0.566

Lightweight document review

dev-review-room

Lamina led
Direct agent
0.566
Lamina agent
0.670
Plan-first agent
0.515

A tiny household list

dev-simple-list

Lamina led
Direct agent
0.614
Lamina agent
0.721
Plan-first agent
0.614

One-sided preference toggle

dev-toggle-preference

Lamina led
Direct agent
0.541
Lamina agent
0.617
Plan-first agent
0.541

Methodology

Three workflows, the same four tasks, and the same 25-minute agent budget per arm.

Direct agentshape_build → verify_fix · 1,500 seconds (25 minutes)
Normal coding agent without Lamina skills. Implements from the founder brief with no plan-first workflow.
Lamina agentlamina_init → lamina_design → implement → fix · 1,500 seconds (25 minutes)
Full Lamina loop: init, design with persona panel, implement, and fix. No separate verify step.
Plan-first agentshape_build → verify_fix · 1,500 seconds (25 minutes)
Normal coding agent using a plan-first workflow. Writes a short plan, then implements. No Lamina skills.
Harness
cursor-cli
Model
composer-2.5
Reruns
3 per task and arm
Scoring
RewardKit · openai/gpt-5.5

RewardKit scored /app/app.mjs and /app/ui.mjs. Lamina process notes were not scored.

Scope

Current release

Four thin vertical slices, greenfield projects only, using Cursor CLI with Composer 2.5. Results apply to this corpus and setup—not every software project.

Next

Larger applications, brownfield projects, and more model and harness combinations.

Evidence

Inspect the packaged README, task definitions, and public jobs on Harbor.

Open Harbor dataset (opens in new tab)