Subnet 66 turns software engineering into a tournament: coding agents fight to fix real GitHub bugs, and only the reigning champion gets paid.
//What is Ninja
Ninja (Subnet 66) runs a continuous contest between AI coding agents. Miners build small programs that read a real software bug and write a patch to fix it, and the subnet pits those programs against each other to find the one that fixes problems best. The single strongest agent at any moment is the "king" and keeps the title until a challenger beats it.
The simple version: It's like a chess ladder, except the players are AI programmers and every match is a real GitHub bug that needs fixing.
Centralized equivalent: Think the SWE-bench leaderboard, or a coding assistant like Cursor or Devin, but run as an open, always-on competition where the winning agent is decided on-chain instead of by a vendor's private scoreboard.
How it works:
Miners build a solver agent: a small Python program that takes a code repository plus an issue description and returns a patch.
Validators generate tasks from real commits, run each agent inside a sandboxed Docker container, and score its patch against the known-good fix using a changed-line similarity check and an LLM judge.
//Why This Matters
The problem it solves:
Measuring whether an AI can actually fix software is hard, and easy to game with cherry-picked demos. Ninja grounds it in real bugs mined from real repositories and settles it through head-to-head duels, so an agent has to genuinely fix issues to win.
The opportunity: Software engineering agents are one of the most closely watched corners of AI right now. A neutral, always-on arena that ranks them on real tasks is useful signal in a field where most numbers come from one-off benchmark runs.
The Bittensor advantage: The competition runs continuously and pays out automatically. Anyone can enter an agent, and the ranking is settled on-chain rather than locked inside one company's leaderboard.
Traction signals: The miner starter harness has been forked 91 times, the validator codebase logged its most recent commit on 2026-06-24, and the team has launched a workspace called Katana alongside an active @Ninja_Subnet account. The active miner field is still small, which fits a format where one champion holds the title at a time.
//Full Analysis
Category: Code Generation and Development Tools | Centralized Competitor: SWE-bench leaderboards, Cursor, Devin
AI coding agents have moved from autocomplete to systems that open pull requests, and benchmarks like SWE-bench have become the scoreboard everyone cites. Ninja's bet is that the scoreboard itself should be live, adversarial, and decentralized rather than a static dataset run once and published.
Mechanism:
Miners write a solver: a Python agent.py that exposes a solve(repo_path, issue, model, api_base, api_key) function and returns a patch. In production, miners do not open public pull requests. They submit the agent privately to the validator's API, signed by their . Before a submission is queued, the validator runs three checks: Agent Smoke (it must compile and pass a pyflakes lint), a Submission Scope Guard (it cannot break the solve contract or hardcode its own model, provider, or sampling settings), and an OpenRouter Submission Judge that reviews the diff. The validator supplies the model, endpoint, and key, so agents compete on prompting and tool-use strategy rather than on who can afford the bigger model.
Scoring is king-of-the-hill. The validator keeps two fixed pools of 50 tasks each, mined from real commits. A challenger first duels the current king on the primary pool, and if it wins it has to win again on a separate retest pool before it can take the crown. That two-pool rule guards against an agent that got lucky on a single task set. When a challenger is promoted, the validator publishes the winning agent.py to the public base repo and assigns all of its weight to the winning hotkey on the next weight-setting . Reward concentrates on whoever currently holds the title.
The infrastructure is specific. According to the repository, solver, task-generation, and evaluation traffic runs on a self-hosted Qwen3-32B endpoint, while the diff judge and the submission judge run z-ai/glm-5.2 through OpenRouter at temperature 0. Crowned kings are also benchmarked against external suites, SWE-bench Verified, SWE-rebench, DeepSWE, and Terminal-Bench, measured against a fixed baseline agent. That external benchmarking is where the "distilling software agents" framing comes from: the tournament is meant to surface and harden a single best agent rather than reward a crowd.
One note on identity. The article we published for netuid 66 in April covered a project called Tau. The slot was re-registered on 2026-04-27 and now carries the on-chain identity "ninja" at ninja66.ai, built on the same unarbos/tau codebase. That development repo, created on 2026-03-31, has 324 commits on its main branch from 8 contributors, with the most recent on 2026-06-24 (verified by a live GitHub check). On-chain at , the subnet holds roughly 0.75% of network , with an alpha price of 0.00903 TAO, a near 46,581 TAO, and a pool depth around 11,377 TAO. Price was up about 13% over the prior week and down about 12% over the prior month, and the subnet showed 5 active miners.
//Risk Factors
These factors move fast; captured at publishing date
Winner-take-all rewards: By design, validator weight concentrates on the single reigning king, so miner emissions flow to one hotkey at a time. That sharpens the competition, but a miner earns little until its agent actually takes the crown.
Crowded category: Software engineering agents are contested ground, both inside Bittensor and outside it, where Cursor, Devin, GitHub Copilot, and the broader SWE-bench ecosystem are well funded and moving fast. Ninja competes for the same mindshare.
: The pool is modest, around 11,377 TAO of depth with roughly 67 TAO of net volume over the past week, so larger positions can move the alpha price and face .
Young on-chain identity: The slot was re-registered and rebranded in late April 2026. The codebase has continuity, but the on-chain identity is new, and its 4-month immunity from deregistration lapses around late August 2026, after which a low price would expose it to the usual deregistration mechanics.
Into the next one.
Keep exploring
Other research from the same neighborhood of the network.
