Table of Contents
Ridges AI, the software-engineering agent network on Bittensor Subnet 62, has completed its Linting competition with a winning agent that scored 90% at an average cost of $0.03 per run. The starting baseline scored 38% and cost $0.21, putting the winner 52 percentage points higher at roughly 86% lower cost.
The winning submission, wingsuit v3, emerged from 821 submissions reported in Ridges' final results. Fifteen agents were approved for emissions, the subnet's token rewards.
Linting involves checking code against rules intended to catch errors and enforce coding conventions. Ridges organizes coding-agent development into specialized competitions called Niches, where participants, known as miners, compete to improve performance and reduce the expense of running agents.
The Competition 27 leaderboard shows how those two objectives shaped the contest. Early in the competition, smCat v1 reached 68%. Its successor, cowrie v0, raised the score to 70% while cutting average cost from about $0.12 to $0.01. The leaderboard records an 88.7% cost reduction for that improvement.
Later entrants pushed scores higher: butterfly v1 reached 82%, grouse v1 reached 86%, and IAMHERE v5 reached 90%. Wingsuit v3 matched IAMHERE's score but took the lead by using 19.7% less inference spending, the cost of the model calls used to perform the task. Its average runtime was 3 minutes and 43 seconds.
The results follow Ridges' earlier introduction of cost controls in Competition 22, when it imposed a maximum cost of $0.29 per problem solved. That competition produced nine emissions-approved agents from 419 created.
For its next competition, Ridges is introducing miner-set baselines through a benchmark phase. Qualifying agents will compete to establish the initial leader before emissions begin. Subsequent entrants will then need to improve on that leader, with the default rules allowing qualification through either a relative score increase of at least 3% or a cost reduction of at least 6% while matching or exceeding its score.
