Kaboo

Kaboo is an open-source project inside ByteDance that analyzes token consumption.

The most interesting thing about it, though, isn't what it does but how it was made: with a new way of working together.

There's no relay of people here, no "file a requirement, schedule it, build it, test it, ship it." Whoever raises a problem designs the fix, runs the tests and deploys it to production themselves. Each person owns their part, end to end.

Quality rests on two safeguards:

1. Before release, an agent checks the change for problems.
2. After release, anything that goes wrong can be rolled back quickly.

It works because the infrastructure is in place. We build on mew, a project-management platform that builds agents into the software-development process, and it is connected to Lark.

The road from idea to implementation is already paved. As long as you can say clearly what you want, you can build it and ship it yourself.

Working this way, I owned the Token Efficiency Leaderboard and the efficiency-analysis module.

I raised the problem. Everyone looked at where agents failed; nobody looked at tasks that succeeded but were wasteful along the way.

I designed the fix:

• I defined criteria for five kinds of waste.
• I designed a diagnosis that uses only metadata and never touches content.
• I normalized cost across different models with Effective Tokens.

Then I tested it and shipped it myself. It now covers more than 10,000 internal users.

This gave me a very concrete picture of how collaboration works in the AI era. Once the infrastructure flattens the cost of implementation, the bottleneck becomes whether you can think a problem through and say it clearly.

A person's value lies more in judgment than in execution.