You're guessing
It looks fine out of the box. But is it actually good at your work? You can't tell.
Tries new versions. Reviewers check them, blind.
Illustration of Spark Loops testing your agent with separate AI reviewers.
Spark Loops tests and improves your AI agent: LLM models from different AI Labs review it blind, your agent mutates and keeps getting better, until it masters the workflow.
Proven, not guessed.
Why you need it
Different AI models review its work and give you an honest answer. Then it keeps improving, on any workflow.
It looks fine out of the box. But is it actually good at your work? You can't tell.
Several different AI models review its work, so you see how good it really is.
It keeps getting better on the fly, on any workflow, without you watching over it.
How it works
Every round ends with upgrade ready, nothing changed, or not proven. The next round starts from your current best.
Spark reads your instructions and examples, then writes a handful of new versions. Plain text you can read.
+ If the budget is missing, ask before you decide.
Every version works through the same real examples: the good ones, the bad ones and the tricky cases.
Same examples, every version.
AI reviewers from different companies compare answers labeled A and B. Your agent never grades itself.
Answer A or answer B? Names hidden.
The leader faces fresh examples it has never seen. A win that came from memorizing fails here.
Fresh examples, never seen in training.
A version that proves better is one click from work. If nothing does, you keep what you had. You can always roll back.
Sample result
The change in plain text, who preferred it, and how it did on examples it never saw.
The change
-Rate each lead as hot, warm or cold.
+Say fit or not a fit, with one reason.
+If the budget is missing, ask before you decide.
Use cases
Works best on a clear, repeated task where you can show good and bad examples.
Sorts inbound leads into fit or not a fit, with a reason.
Drafts replies to common questions in your voice.
Checks posts against your rules: claims, tone, banned phrases.
Answers your users inside the app you built, within your rules.
Checks a change against your project's rules before you merge.
Sorts bug reports by area and urgency, and asks for missing steps.
Routes tickets and drafts a first reply.
Drafts outreach and follow-ups.
Starts onboarding only after a positive reply or a booked meeting.
Accepts or rejects applications, requests or leads under your rules.
Drafts customer or partner messages within strict rules.
Checks documents against a scoring guide.
Sorts your email into reply now, reply later or ignore.
Rewrites your drafts so they sound like you.
Turns long articles into the points you care about.
Examples of what you could train, not customer stories.
All use casesWhy trust it
AI reviewers from different companies compare versions without knowing which is which.
A hidden set of fresh examples decides whether the win is real.
Training runs on your machine. Your examples go to the model providers you connect.
Set a budget and training stops there. You pay providers directly, no markup.
Compare
| Question | Hand-tweaking prompts | Code eval tools | Fine-tuning | Spark Loops |
|---|---|---|---|---|
| What changes | Your prompt, by hand | Nothing. They measure. | The model itself | Your agent's instructions |
| Needs code | No | Yes | Usually | No |
| Blind outside reviewers | No | If you build them | No | Built in |
| Hidden test | No | If you set one up | If you set one up | Built in |
| Readable change | Yes | Nothing changes | No | Yes, plain text |
| Switch models later | Yes, then retune by hand | Yes | No, it is tied to one model | Yes |
No published benchmark results for Spark Loops yet. This compares approach, not measured speed or quality.
Where it's going
Now
One agent, one job, made better on your computer.
Next
Reusable chips for common jobs, and training you can plan from a Grok Bot chat.
Later
Keeps training in the background, so an upgrade can be ready by morning.
Vision
Sparks share what worked, so each one can start from the best version anyone has found.
Holders get paths into the vibe/vibe ecosystem.
No. It improves your agent's instructions. The model stays the same, and it is not fine-tuning.
No. Pick one clear job, add real examples, and read the results in the app.
You pay your model providers directly, and training stops at the spending limit you set.
Training runs on your computer. Your examples go to the model providers you connect.
No. It trains and improves on its own, inside the spending limit you set.
Mac and Windows, early access. Contact us on X