Ridges is tightening up ahead of its March 5 open beta launch, and the team aims to align the subnet with the product, or nothing else matters.
During this interview with Alchemist from Subnet Summer, Cameron (team member at Latent Holdings) and Shaq (owner of Ridges SN62) made it clear that the recent delay was about making sure evaluation metrics actually reflect real-world usage.
Here’s the interview:
Right now, agents get generous evaluation windows and are judged mostly on benchmark outputs. That’s not how production software works. So they’re adding linting, standard PR checks, inference cost tracking, and tighter evaluation logic. The goal is to ensure what wins the subnet competition should also be what users want to run in their repos.
The long-term vision goes beyond coding. Ridges is positioning itself as a generalized agentic competition framework. Today, it’s development PRs. Tomorrow, it could be smart contract auditing, personal assistance, or other benchmarkable agent tasks. The architecture is modular enough to spin up sub-competitions under one subnet umbrella. If you can define the benchmark and weight it properly, Ridges can likely host it. That’s the thesis.
A major focus post-beta is speeding up evaluations. Right now, slow eval cycles limit how quickly miners can iterate and how many teams can compete daily. Faster elimination logic means more competition, and more competition means better agents. They’re also shifting toward a bring-your-own-API-key model. Instead of Ridges proxying inference and estimating costs upfront, miners will choose whitelisted providers and pay directly. That forces optimization. When every call costs you personally, you write tighter, more efficient agents. That efficiency translates directly to better product performance.
They’re also integrating newer benchmark suites like Harbor and moving toward synthetic evaluations. Rather than relying on static public datasets that invite overfitting, they plan to pull real-world successful PRs, filter and classify them by difficulty, and convert them into fresh problem sets. That keeps the competition dynamic and closer to production reality.
On the product side, the demo showed a working GitHub integration. You open an issue, add a Ridges label, and the agent reads the context, drafts a PR, summarizes its plan, and submits code. The demo displayed by the team shows that it’s already writing and modifying production code. Pricing for the beta is set at $29 per month, with around 100 quota-based requests. They’re considering discounts for paying in alpha, potentially reducing sell pressure or supporting burns, though token mechanics aren’t finalized. The focus right now isn’t financial engineering, but ensuring that the best product is shipped.
Stability has been another major theme. The subnet has faced consistent DDoS attempts. Instead of ignoring it, they hardened the infrastructure. The frontend is now separated from the subnet core, and they’re improving validator stability and Kubernetes deployment to reduce costs and increase evaluation throughput. Running a subnet means people will try to break it. The upside is that every exploit attempt forces improvements. In terms of ecosystem strategy, they’re pragmatic about infrastructure. They’ve used Targon and Chutes for inference, dealt with stability and cost trade-offs, and are experimenting with load balancing and OpenRouter despite higher costs. The preference is to use subnets when they’re the best option, not just out of loyalty but because they’re cheaper and competitive.
There’s also an interesting angle around subnet composability. Cameron floated the idea of Ridges agents mining other subnets like Git-based bounty systems by generating PRs automatically. If a subnet rewards accepted pull requests, why can’t an autonomous coding agent compete for those rewards? That’s the kind of recursive abstraction Bittensor makes possible.
For now, Ridges is intentionally under the radar. No big enterprise push yet and no loud marketing blitz. The plan is to polish the product until it can function as a daily driver, then push outward. The partnership between Latent and Ridges is currently seamless, built on prior collaboration and complementary strengths. If they execute, Ridges won’t just be another coding agent. It will be a competitive training ground where the best agents earn their way into production. In Bittensor terms, that’s the right way to build.
Be the first to comment