Request early access
Illustration: an agent on a test floor with three reviewers, each in a different color.

Tries new instructions. Reviewers check them.Illustration. You approve every upgrade.

Illustration of Spark Loops testing new instructions with separate reviewers. You approve every upgrade.

Better agents, without the guesswork.

Test better instructions for your AI agent. You approve every upgrade.

Request early access See how it works

Mac and Windows, early access.
Free during early access.

Spark Loops tests new instructions on your own examples. Review the results. You decide what to keep.

Your agent

Same app, your models.

Spark improves your agent's instructions, not the AI model. Every change is plain text you can read.

Suggests improvements

Spark writes new versions of your agent's instructions.

Checks them, blind

AI reviewers from different companies.

Does the work

Your connected models. You pay the providers directly.

What you need

How it works

Try an idea. Check the result.

You approve every upgrade.

  1. 01

    Try.

    Spark suggests new versions of your agent's instructions.

  2. 02

    Test.

    Each one works through your real examples.

  3. 03

    Review.

    AI reviewers from different companies compare them without knowing which is which. Your agent never grades itself.

  4. 04

    Hidden test.

    A separate test checks fresh examples.

  5. 05

    You decide.

    Read the results and approve a change, or keep what you had.

See the full round

Pick a job

One job, trained well.

Start with one repeat task and real examples. Read the results before you approve a change.

All use cases

Examples of what you could train, not customer stories.

What makes it honest

It never grades itself.

Outside reviewers

Separate AI reviewers check the work without knowing who wrote it.

A test it can't study for

A separate test checks how the new instructions handle fresh examples.

You decide what stays

Read the results before you approve an upgrade. You can keep what you had.

Runs on your computer

Your gym. Your machine.

A desktop app. Training runs on your computer, with your own model keys. You pay providers directly.

Mac and Windows, early access.

Platforms and install
  • The gymTraining and testing happen here, on your computer.
  • Your keysConnect the model providers you choose.
  • Your budgetYou set a spending limit. Training stops there.

Compare

How it compares.

Spark Loops brings instruction testing into a desktop app. Other approaches let you build your own experiments or change the model itself.

Find the approach that fits

Choose what fits

  • Use an app to test instructions on your examples.
  • Use code for experiments you want to build yourself.
  • Spark Loops has no published benchmark results yet.

Built in public

Built with vibe/vibe builders.

A launchpad for AI builders and vibe coders. Get the guide, improve one agent, share what you learn.

Join the builders Getting started guide @spark_coded

Pick one job. Make it better, for real.

Request early access Read the docs

Mac and Windows, early access. Free during early access.