Ridges

SN62 Ridges runs an open tournament for robot software mechanics: builders enter programs that fix broken code, the best fixer earns the prize, and the mechanics are hired out to developers at $9.99 a month for ten repairs.

The commodity, explained from zero

The mechanic is a software engineering agent: a program that reads a description of a bug or a feature request, edits the code, and hands back a patch. Ridges sells it through Ridgeline: connect a GitHub repository, open an issue, and the agent returns a pull request from an isolated machine. "1 credit -> 1 GitHub Issue -> 1 Pull Request." A subscription of $9.99 a month grants 10 credits; alternatively, locking 1,000 of the subnet's tokens earns one credit every three days.

The customer is a developer with a GitHub repository. The roadmap says the subnet launched in late 2024 as AgenTao, that the Ridgeline beta arrived in early 2026, and that in mid 2025 its agents "reached state-of-the-art performance among open-source SWE-bench agents", SWE-bench being a standard test of fixing real GitHub issues. No score accompanies that claim. No usage or customer figure for Ridgeline appears anywhere.

Why it is on Bittensor at all

The docs' argument is against closed agents: "Closed agents may also take the shortcut of hardcoding known solutions. Ridges removes that most basic of cheating strategies structurally." Problems are hidden and randomized, several independent validators must agree, and once an agent stops earning, "the exact agent.py that earned the score is open-sourced". The comparison breaks here: a mechanics' tournament has a sponsor who pays the purse, while the purse here is emission, the subnet's share of newly minted TAO, and the mechanic's tools (the language model calls) are paid for by the builder, at $10 to $20 per evaluation run.

How the work gets done

Miners (the builders) submit one Python file exporting agent_main, which takes a problem statement and returns a patch, paying a burn of about $5 per submission with one submission per hotkey every 12 hours. Validators (the judges) apply the patch and run a hidden test suite: two screening rounds of 20 problems at 45 and 60 percent pass thresholds, then three validators on 50 problems each, with credit only when every validator agrees. An agent earns emission by scoring at least 3 percent above the leader (a leader at 0.50 needs 0.515) or by being 6 percent cheaper at the same score, with a multiplier that doubles after a 6 hour stall and a 14 day half-life on every agent's reward.

How you would know it works

The leaderboard at ridges.ai/agents shows two open competitions dated September 9 and August 18, 2026, but only in a browser; the site returns a Vercel security checkpoint to anything else. No dated benchmark table with numbers exists in the materials.

What is missing

The chain contact hello@ridges.ai appears on no docs page fetched, and the site front page cannot be read by a fetcher. There is no API reference; Ridgeline is a GitHub app and a task board. Support is a FAQ and Discord, sales are final, and there is no status page or SLA. The docs and the chain identity link different Discord invites. Demand for Ridgeline is stated nowhere.

Go deeper

• Docs root, plain account
• How emissions are earned
• Credits and the $9.99 plan

Metaphor: an open tournament for robot software mechanics. Every claim is drawn from the evidence set or the subnet's own materials; "(inferred)" marks a conclusion rather than a quote. Corrections.