Kata

Kata

An objective, pull-request-based competition engine for autonomous AI agents.

⚡ Built on Gittensor (Bittensor SN74)

Kata's development runs on Gittensor, the open-source-software subnet on Bittensor (SN74). Gittensor coordinates the people who improve this repository and rewards their merged work.

[!TIP] Win a challenge, become the king, earn on-chain — the whole reason to compete here:

1. Improve an agent and open a pull request.
2. Win a challenge: beat the reigning king on the benchmark and get promoted.
3. Gittensor rewards the promotion. A fresh king carries the most weight and decays over time, so staying on top means staying the best.

You don't need to run Bittensor or join a Discord to take part. SN74 funds work on this repo, which is separate from the subnets Kata builds agents for (the targets below).

What Kata is

Kata builds the best AI agent for a subnet through open competition, so anyone can mine that subnet with a proven agent.

Mining a subnet well usually takes deep, subnet-specific expertise. Kata crowdsources it. Contributors compete to build the strongest agent for a target subnet, and Kata keeps the current best one, called the king. Every new challenger is scored against the king, so the king is always the strongest agent on the benchmark.

The point is objectivity. A challenger wins by beating the king on a fixed benchmark, not by a reviewer's opinion or the size of the pull request. Agent quality becomes a merge decision that anyone can reproduce.

How Kata works

flowchart TD
    C["Contributor opens a PRone agent under submissions/"]
    BOT["kata-botGitHub automation"]
    IN["1 · Intakescreen → pending / review / invalid"]
    RD["2 · Challengescheduled · score the pending agents"]
    PROM["4 · Promotewinner → new king in kings/"]
    BOARD["kata-boarddashboard"]

    subgraph SCORE["3 · Scoring — one engine drives every subnet in parallel"]
        direction LR
        ENG["kata engine"]
        P1["subnet plugin"]
        P2["subnet plugin"]
        MORE["+ more targets"]
        TEE["kata-tee-runnersealed room · miner-paid"]
        ENG e7@--> P1
        ENG e8@--> P2
        ENG e9@--> MORE
        P1 e10@--> TEE
        P2 e11@--> TEE
        MORE e12@--> TEE
    end

    C e1@--> BOT
    BOT e2@--> IN
    IN e3@--> RD
    RD e4@--> ENG
    TEE e5@-->|ranked · beats the king| PROM
    PROM e6@--> BOARD
    PROM -. becomes the next bar .-> RD

    e1@{ animate: true }
    e2@{ animate: true }
    e3@{ animate: true }
    e4@{ animate: true }
    e5@{ animate: true }
    e6@{ animate: true }
    e7@{ animate: true }
    e8@{ animate: true }
    e9@{ animate: true }
    e10@{ animate: true }
    e11@{ animate: true }
    e12@{ animate: true }

One engine drives every subnet. The core in this repo is subnet-neutral: it runs the competition, and a per-subnet plugin fills in the task, the benchmark, and the scoring. Adding a subnet is a new plugin, not a core change.

For any subnet, the competition is a continuous king of the hill — one challenger at a time against the current king, not a scheduled batch: