You’re already dangerous with these tools.

I run the experiments and hand you the loop, free.

Start with the receipts: Episode 1, the cheap run.

Episode 1 · The cheap run

The cheaper run cost 30% less and shipped nothing that worked.

Clipped names cost $0.44 a run against $0.62, and delivered 0 of 5 working features, twice. Full names, with naming rules in the instructions, delivered 2, then 4 of 5. Here are the receipts.

Episode 1 at a glance

Clipped names

$0.44per run

Best run: 0 of 5 work

  • See all pets: fails
  • Open one pet: fails
  • Add a pet: fails
  • Edit a pet: fails
  • Delete a pet: fails

Full names

$0.62per run

Best run: 4 of 5 work

  • See all pets: works
  • Open one pet: works
  • Add a pet: works
  • Edit a pet: works
  • Delete a pet: fails
The numbers
Episode 1 at a glance: two runs per arm, same task, same model.
ArmMean cost per runMean tokens readWorking features, run 1Working features, run 2
Clipped names$0.43782.38 million0 of 50 of 5
Full names$0.62463.92 million2 of 54 of 5

Free · MIT licensed

The plan, execute, reflect kit.

The loop I use to run every experiment, packaged so you can run it on your own project. It works with Claude Code and Codex, installs in about 10 minutes, and comes with four sample business tasks.

Get the free kit

The loop ships real products

Built with the same loop, and live today.

  • Scribe ProA native macOS dictation app that works on your device.
  • HeatPilotA free, notarized menu-bar fan controller for Apple Silicon.
  • Flipping Command CenterA live web app that computes max bids from sold comps.

I spent roughly five years as a security engineer on infrastructure that helps protect U.S. elections, in SOC 2 Type II and ISO 27001 environments. The loop carries that rigor.

More about AgentC →

Tell me what you’re building.

I will point you to the right next step, or build you a small working version so you can see it for yourself.

or build it together on a call