How it works
Give your agent better instructions.
Test new instructions. Review the results. You approve every upgrade.
Start with one job.
Choose a repeated job and gather real examples of good and bad work. Your agent's instructions, examples and rules live together.
Spark changes the instructions. The underlying model stays the same. Each change is text you can read.
Create your first agentFrom an idea to your decision.
-
01 · Try
Try
Spark suggests new versions.
-
02 · Test
Test
See how the new instructions handle your examples.
-
03 · Review
Review
Outside reviewers compare them, blind.
-
04 · Hidden test
Hidden test
A separate test checks fresh examples.
-
05 · You decide
You decide
Read the results and approve a change, or keep what you had.
Separate reviewers check the work.
AI reviewers compare the versions without knowing who wrote them. Your agent does not grade itself.
Read the change before you use it.
You review the new instructions and results, then approve the upgrade. You can keep the current version instead.
Put results in context.
Look at the examples and reviewer feedback alongside scores. Different tests can give different results.
Reading resultsProgress can stall.
Review the job and examples before running more training. Spark also offers further experiments when progress stalls.
When progress stallsLearn from real work too.
Record real outcomes, such as how leads turned out. They help you understand what a test alone cannot show.
Set a spending limit.
You pay your model providers directly. The app shows estimated costs and stops training at the spending limit you set.
Running a training runKnow where data goes.
Training runs on your computer. Prompts and examples go to the model providers you connect.
Read the draft privacy pageComing soon.
Grok Bot chat training and connections to outside helpers are coming soon. For now, start in the desktop app.
Pick one job. Make it better, for real.
Mac and Windows, early access. Free during early access.