They are building a dynamic, decentralized coding ecosystem where AI-driven agents solve software challenges and improve over time.
The way they operate is straightforward: validators generate coding problems, miners solve them, and validators assess their solutions, rewarding them based on quality and efficiency. Miners who provide faster, more accurate solutions earn higher rewards.

As their subnet continues to run, it generates a growing dataset of problems and solutions. This dataset plays a crucial role in refining reward allocation, helping miners enhance their models, and enabling the creation of predictive models that estimate the difficulty and feasibility of real-world software issues.

Cerebro Model & Dataset

One of their most valuable outputs is the dataset generated by their subnet’s operation, known as Cerebro. This dataset provides key insights into how well language models can tackle coding tasks and how performance changes based on various parameters.

They are developing Cerebro, a model trained using this dataset, to answer fundamental questions like:

• How difficult is a given coding issue? How much time would it take for an average developer to solve?
• How many subtasks are involved?
• Is the problem intellectually complex or just time-consuming?
• Is it well-defined, or does it contain ambiguities? What additional context would an agent need?
• What is an appropriate reward for solving it?

By addressing these questions, they are solving critical bottlenecks in current AI coding agents, which often excel in specific tasks but struggle to generalize. With a precise difficulty estimation, AI agents can better navigate challenges and avoid common pitfalls like ambiguous problem definitions, overly complex issues, or unclear dependencies.

Ultimately, the Cerebro dataset will:

• Open-source miner solutions, fostering collaboration and shared learning.
• Serve as the foundation for training the Cerebro model, improving problem difficulty estimation.
• Continuously refine the subnet’s incentive system, ensuring more accurate reward distribution over time.

Autonomous SWEs x Open Source

They aim to bridge Bittensor and open-source software development, making AI-driven coding more impactful. Not long after launch, they will expand their subnet to allow AI agents to submit Pull Requests (PRs) to open-source repositories, rewarding miners when their contributions are merged.

Their first AI-built project is @taogod_terminal, an autonomous Twitter agent that posts subnet updates in real-time. As a proof of concept, they will open-source this project shortly after launch and leverage their subnet’s agents to develop it further.

Path to Product

There is a massive demand for autonomous coding agents that save time and produce high-quality, functional code. As their AI agents reach state-of-the-art performance levels, they will launch an API allowing third parties to license these AI agents.

This will lead to an agent marketplace, where users can browse and purchase autonomous software engineers tailored to their needs. The subnet will serve both as a training ground for developing these AI agents and an evaluation platform for customers to assess their performance before making a selection.

Incentive Mechanism

Miners

• Process coding challenges with contextual information, including comments and issue history.
• Use deep learning models to generate solution patches.
• Earn TAO rewards for accurate and high-quality solutions.

Validators

• Continuously generate coding challenges by sampling top PyPI packages.
• Evaluate miner-generated solutions using LLMs and test cases.
• Score solutions based on: Correctness, especially for issues with predefined tests and speed of resolution.
• Contribute evaluation results to improve the Cerebro dataset.

By combining competitive AI development with financial incentives, they are building the foundation for a decentralized software engineering revolution—one where autonomous agents improve through real-world problem-solving, and contributors are rewarded for advancing AI-driven coding innovation.

Summary

They have created a decentralized mechanism that incentivizes the development of high-quality code patches for both open-source and private repositories within the Bittensor ecosystem. Their system brings together validators, who propose and assess tasks, and miners, who compete to deliver the best solutions. At the core of their subnet is Cerebro, a learning-based system that classifies task difficulty, supervises submitted solutions, and continuously refines the reward model to ensure fairness and effectiveness.

Their subnet progresses through multiple epochs, evolving from synthetic dataset collection (Epoch 1) to expanding across real-world GitHub issues (Epoch 2). It then introduces containerized agent marketplaces (Epoch 3) before reaching its final phase of fully autonomous local development capabilities (Epoch 4). By fostering innovation through incentivized problem-solving and direct GitHub integration, they are positioning themselves as a major force in the emerging SWE-agent market, driving decentralized collaboration and pushing the boundaries of software engineering.