Count the whole task.
A cheaper model call means little if it takes eleven attempts to finish the job. Count the retries, the waiting, and the person who had to step in. The outcome is what matters.
About Caveman
We’re here to make AI agents do better work, with less waste. Less cost. Less waiting. Fewer failures.
It started with a student’s AI billA student’s budget. An unexpected line item.
How we got here
A very personal
waste problem.
I was a student building with AI. Then my AI bill grew to five times what I spent on groceries.
I wanted to keep building. So I started looking at what I was paying for: repeated context, long answers, and work the agent had already done. I made a small fix and put it on GitHub.
Overnight, Caveman reached #1 on Hacker News. A prompt that made agents talk like cavemen had found a problem a lot of people recognized.
Then enterprises started getting in touch. I worked with them one on one, helping untangle their own AI waste. The same question kept coming back: how do we get useful work out of these systems without spending so much money and time getting there?
That work became the platform. The problem had grown from a student’s budget to entire agent systems. The reason for building stayed the same.
The open-source project that started itThe mission
More work
worth doing.
AI makes more possible. Waste puts a limit on how much of it we can use.
We’re building Caveman to move that limit. To help agents finish the job with less money, less waiting, fewer failures, and less human intervention.
The goal is continuous improvement: understand where resources go, find a better way, prove it works, and put it into practice.
What we optimize for
Cost per successful
agent task.
With quality and reliability held to the standard the work requires.
Total cost, including every attemptSuccessfully completed tasks
One system. A continuous loop.
Explore the platform- 01
Find the waste
Understand what happened across the whole task.
- 02
Try a better way
Turn that evidence into a change worth testing.
- 03
Check the outcome
Measure cost, speed, quality, and reliability together.
- 04
Put it to work
Roll out proven improvements. Keep learning from the next run.
How we build
The Caveman way.
Make better prove itself.
An improvement has to survive contact with real work. Test the change against the task. Check the quality. Measure the total cost. Keep the results that hold up, and learn from the ones that don’t.
Earn the right to act.
The goal is less work for the people running agents. Automation should earn their trust through evidence, clear limits, and changes they can review and reverse.
The ambition got bigger.
The idea stayed simple.
why use many token
when few token do trick.
Keep building.
Let’s make more
of what AI can do.
If you’re spending too much time or money getting agents to work, we’d like to hear your story.
Talk to us contact@caveman.ac