Ridges AI is a decentralized artificial intelligence project that operates as Subnet 62 on the Bittensor network. The project, formerly known as Agentao, is focused on creating a marketplace of autonomous software engineering agents designed to perform end-to-end software development tasks. ^{[1] [2]}

Overview

Ridges AI aims to build a decentralized network where AI agents can autonomously solve complex software engineering problems. The project's core thesis is that the multifaceted role of a human software engineer can be deconstructed into a series of smaller, discrete tasks. By training specialized AI agents to master each of these individual tasks, such as fixing code regressions, writing unit tests, or resolving GitHub issues, the platform can aggregate their outputs to deliver comprehensive, end-to-end solutions. This approach is intended to challenge the dominance of large, centralized AI corporations by creating a self-sustaining, incentivized ecosystem for AI-driven software development. ^{[3] [1]}

The platform operates on the Bittensor network, leveraging its crypto-economic incentive mechanism to foster competition among participants. In this system, network participants known as "miners" operate AI agents that compete to complete software engineering tasks. Other participants, called "validators," are responsible for evaluating the quality and efficiency of the solutions provided by the miners. Miners who produce the highest-quality work are rewarded with TAO, Bittensor's native token. This competitive framework is designed to continuously improve the capabilities of the agents on the network. The project's ultimate goal is to create a system where real-world software development needs can be met by autonomous agents, with project revenue eventually sustaining the network's incentive emissions. ^{[3] [1]}

A key representative for the project, Shakeel, has stated that the long-term vision is for the product's real-world usage to dictate network rewards. In an interview, he explained, "The product itself will decide who earns emissions," signaling a strategic shift from synthetic benchmarks to a model where rewards are based on user engagement and the acceptance of AI-generated code contributions in real-world applications. ^[1]

History

The project, originally named Agentao, was rebranded as Ridges AI and established as Subnet 62 on the Bittensor network. Its development roadmap is structured in distinct phases, or "epochs," beginning with the collection of synthetic datasets to train the initial agents. Subsequent epochs focus on expanding the agents' capabilities to solve real-world GitHub issues, introducing containerized agent marketplaces, and ultimately achieving fully autonomous local development capabilities. ^[1]

During the first half of 2025, Ridges AI achieved several key milestones. In early Q2, the team introduced a new incentive mechanism for tasks related to continuous integration (CI) regression and code generation. They also released performance benchmarks against the SWE-Bench standard and opened platform access to Subnet 62 miners. This was followed by the public launch of its API and a leaderboard to track the performance of top-performing miners. Later in Q2, the platform expanded access to all Bittensor miners, added new task types, and launched a detailed performance dashboard. ^[1]

In October 2025, the team announced a full platform rewrite aimed at improving stability, enabling parallel evaluations of agent performance, and implementing a dual-sandbox environment for testing. The project also planned the launch of its first major product, Ridges V1, a Cursor/VS Code extension, for October 30, 2025. This period also saw the project recover from a performance dip on the SWEBench Polyglot benchmark, which the team cited as evidence of their ability to iterate and improve the system rapidly. ^[1]

Technology

Ridges AI is built on the Bittensor network and its codebase is primarily written in Python, with PostgreSQL used for its database and Docker for containerization. The architecture is designed to be model-agnostic, allowing it to function as a "thick agent layer" that can integrate and orchestrate various large language models to create complete coding solutions. The system is also designed for composability with other Bittensor subnets, such as Chutes and Targon, to enhance functionality and reduce operational costs. ^{[4] [1]}

Architecture and Mechanism

The Ridges AI ecosystem operates on a competitive, decentralized model involving two primary roles: validators and miners.

• Validators are responsible for generating software engineering challenges and evaluating the solutions submitted by miners. Initially, these challenges were created by sampling from top PyPI packages, with plans to expand to real-world GitHub issues. Validators use a combination of large language models and test cases to score submissions based on correctness and speed.
• Miners are the autonomous AI agents that compete to solve these challenges. They process the tasks, generate solution patches using deep learning models, and submit their work for evaluation. Miners who consistently provide high-quality and efficient solutions earn TAO rewards, creating a financial incentive for continuous improvement.

This mechanism is designed to generate a valuable dataset of coding problems and their corresponding AI-generated solutions, which is then used to further train and refine the models on the network. ^{[3] [1]}

Core Components

The platform is composed of several key technical components that work together to manage the network and its tasks.

• Cerebro Model & Dataset: This is a central component generated from the ongoing operations of the subnet. The dataset is a continuously growing collection of coding problems and their verified solutions. The Cerebro model is trained on this data to perform several critical functions, including estimating the difficulty of new tasks, identifying ambiguities in problem descriptions, determining appropriate rewards for solutions, and refining the subnet's overall incentive mechanism.
• Problem Routing Protocol: This component acts as an orchestrator model. It is designed to deconstruct large, complex software projects into smaller, manageable tasks such as code generation, bug fixing, or CI regression detection. It then routes these discrete tasks to the specialized agents on the network that are best suited to solve them.
• Inference Gateway: A service that handles and routes AI inference requests, acting as a bridge between user queries and the various AI models operating on the network.
• Cave: A local developer dashboard designed to integrate with the Ridges platform, simplifying the development and management process for network participants.

These components are visible in the project's open-source repository, which also includes modules for the agent, validator, evaluator, and API. ^{[4] [3] [1]}

Tokenomics

The native token of the Ridges AI subnet is SN62. It operates on the Bittensor network, and its primary trading pair is SN62/TAO on the Subnet Tokens exchange. The tokenomics are designed with a maximum supply mirroring that of Bitcoin. ^[2]

• Ticker: SN62
• Network: Bittensor
• Max Supply: 21,000,000 SN62
• Total Supply (self-reported): 1,260,000 SN62
• Circulating Supply (self-reported): 1,260,000 SN62

As of late 2025, the project's self-reported market capitalization was approximately $35.42 million, with a fully diluted valuation of around $587.34 million. ^[2]

Products and Go-to-Market Strategy

The initial go-to-market strategy for Ridges AI was to offer its services as a product to other miners within the Bittensor ecosystem. The value proposition was to provide smaller or solo miners with a "suite of software engineers" on demand, allowing them to offload software maintenance tasks and focus their resources on optimizing their primary mining operations. ^[3]

As the project matured, its product offerings expanded to target a broader market, including individual developers and open-source projects.

• API: An Application Programming Interface that allows third parties to license and integrate the autonomous AI agents into their own workflows. The API was in private beta before its public launch in 2025.
• Agent Marketplace: A planned platform where users can browse, select, and purchase the services of autonomous software engineering agents tailored to specific needs.
• Ridges V1 IDE Extension: A Cursor/VS Code extension designed to bring the capabilities of the AI agents directly into a developer's integrated development environment. The planned pricing was approximately $12 per month, with an optional tier at around $8 per month for users who agree to share data to help improve the models.
• Open Source Integration: The project aims to enable its AI agents to submit pull requests directly to open-source repositories. Miners would be rewarded when their AI-generated contributions are reviewed and merged by human maintainers. The team built a proof-of-concept tool, @taogod_terminal, as an early demonstration of this capability.

The project's financial strategy aims for revenue generated from these products to exceed the cost of network emissions by January 2026, with profits being reinvested into further product development and growth. ^[1]

Team and Funding

The development of Ridges AI is managed by a team of contributors visible on the project's public GitHub repository. While formal roles are not specified, several core contributors include aaron-ridges, adamridges, alex-ridges, stephen-ridges, and omar-ridges. Shakeel ("Shak"), whose GitHub handle is hobbleabbas, is a key public representative for the project and has participated in interviews detailing its vision and progress. ^{[4] [1]}

In 2025, Ridges AI received a $300,000 investment from DSV Fund. The investment was disclosed during an interview in August 2025 with DSV Fund partners Siam Kidd and Mark Creaser on their "Revenue Search" series. ^[1]

Controversies and Challenges

In late September 2025, the Ridges AI team identified and addressed a spam attack on its network. According to the team, a coordinated group was exploiting the system to hinder competition. In response, the project implemented a significant policy change: miner-generated code would only be made open source after it had been successfully evaluated and verified. This countermeasure was designed to prevent malicious actors from simply copying and resubmitting valid code from legitimate miners to illegitimately claim rewards. ^[1]

The project also faced performance challenges. The subnet's performance on a mixed-set benchmark that included SWEBench Polyglot experienced a significant drop from a high of 88% to around 17-18%. However, following adjustments and iterations by the team, performance rebounded to approximately 41% by early October 2025. A project representative cited this rapid recovery as a demonstration of the team's agility and the network's capacity for self-correction. ^[1]