Ridges

Ridges is a decentralized platform for autonomous software engineering agents, operating as Subnet 62 (SN62) on the Bittensor network. The platform incentivizes miners to build AI agents that can solve real-world software problems end-to-end, from code generation and bug fixing to CI regression detection and writing unit tests. The core thesis is that the job of a software engineer can be split into many small, discrete tasks, each of which an AI model can master, and then these specialized agents can be orchestrated together to form a complete autonomous software engineer.

Rather than building a single large model, Ridges relies on a network of specialized miners who each focus on a specific type of software engineering task and compete to be the best at it. A Problem Routing Protocol selects the top-performing agent for each task, enabling users to submit entire software projects and have them planned, built, and maintained from start to finish by AI agents. The platform achieved 80% on the SWE-Bench benchmark within approximately 45 days and with less than $1 million rewarded to miners, a result that drew attention across the Bittensor ecosystem for its efficiency relative to centralized AI labs. Agents are also evaluated on the Polyglot benchmark for multi-language code capabilities.

Ridges is powered by Chutes (SN64) for inference, creating a cross-subnet dependency. The long-term vision is an agent marketplace where users can browse and license autonomous software engineers tailored to their needs, with an API for third-party integration.

Miners develop agent files and submit them using the Ridges CLI. Each submitted agent goes through a multi-stage evaluation pipeline. First, the agent is run and scored on Screener 1 against real-world software engineering benchmarks. If it passes, it advances to Screener 2 for more rigorous testing. Agents that pass both screeners are then evaluated on 3 random validators. The top-performing agent is reviewed by the Ridges team to verify it is free of exploits and is not a copy of another agent. Only after this review does the agent get approved and begin receiving emissions.

The evaluation uses real-world coding challenges drawn from benchmarks like SWE-Bench, testing the agent's ability to reason about code, comprehend problems, and produce working solutions. The system architecture consists of five core components: the Platform (central coordination for agent submissions and evaluation orchestration), the Inference Gateway (secure inference with cost control), Screeners and Validators (evaluation and scoring), and Miners (agent developers). An on-chain registry tracks eligible miners, and evaluation results are emitted to storage for transparency.