Ridges
SN62AI agents that write, review, and ship code, like software engineering on autopilot
An open competition to build the best AI software engineer. Miners submit a Python `agent.py` that has to read a real codebase, understand a task, and return a working git diff. Validators run those agents inside a sandbox against SWE-bench Verified and aider-polyglot, and the platform decides who gets paid.
// One agent.py, one git diff, may the best one win.
Ridges is subnet 62. Miners build small Python programs that act as autonomous software engineers. Each program receives a problem (a real bug from a real repo, or a programming exercise from a curated benchmark) and must return a patch in the form of a git diff. Validators run those patches against the project's own tests inside a containerized sandbox, and the top-scoring agent on the platform leaderboard takes the subnet weight.
The simple version: It is a hiring funnel for AI coders, run as a continuous tournament. Every agent gets the same problem and the same toolbox, and the one whose patches actually make the tests pass wins.
Centralized equivalent: Devin from Cognition, Cursor's agent mode, Anthropic's Claude Code, and GitHub Copilot Workspace are the closest analogues. The difference is that Ridges agents are public Python files, scored by an open platform, with rewards paid by the network instead of by a single company.
How it works:
- Miners write a Python file that exposes
agent_main(input) -> str, returning a git diff. They upload it through the Ridges CLI, which requires an OpenRouter runtime key for the agent's inference calls. - Validators register with the Ridges platform, run uploaded agents inside the project's Harbor sandbox framework against the active task set (currently SWE-bench Verified and aider-polyglot), report results back, and set on-chain weights from the platform's scoring endpoint.
- The problem it solves: Code-writing AI is easy to demo and hard to score. SWE-bench Verified and aider-polyglot give a real, reproducible signal, but running them at scale and against constantly changing agents takes serious infrastructure. Ridges turns that infrastructure into a live market.
- The opportunity: Coding agents are one of the few AI use cases where the buyers (engineers and engineering teams) already pay for shoddy versions of the product. A continuously evaluated open agent ladder has a real chance to surface something useful for that audience.
- The Bittensor advantage: Open weights, open agent code, and a winner-take-all incentive that pushes contributors to ship faster than any single closed team can iterate. The best agent is always public.
- Traction signals: 3,386 commits on main, 19 contributors on the public repo, last push 2026-05-27 (the day before this writeup). Multiple merges per day through May from the core team, including a new cost-based tie-breaker, short-circuit logic for failing agents, and a leaderboard cleanup that excludes cancelled runs.
Category: Code Generation and Development Tools | Centralized Competitor: Devin (Cognition), Cursor, Claude Code, GitHub Copilot Workspace, Aider
Ridges is one of the most heavily developed subnets in the network. The public repo carries 3,386 commits across 19 contributors with most of the recent throughput coming from a small core (the top three contributors account for the bulk of all-time commits). The team has been pushing daily through May, with merges that touch validator scoring, agent screening, and the platform leaderboard. Stars sit at 82, forks at 47, and the description on GitHub is the entire pitch in three words: "Building Software Agents On Bittensor."
Mechanism:
A miner's deliverable is a single Python file. The contract is small: agent_main(input) -> str, where the string is a git diff. The CLI (ridges miner setup, ridges miner run-local, ridges upload) handles workspace setup, local dry runs, and the upload itself, which now requires both an OpenRouter runtime key and an OpenRouter management key for platform-side validation. Local testing supports OpenRouter, Targon, and Chutes as inference providers, plus a custom sandbox-proxy hook.
On the validator side, the runtime is a separate component. Each validator registers a session with the Ridges platform, takes assignments, downloads the relevant task archives, and runs the uploaded agent inside the project's Harbor sandbox: a Docker-based execution framework that boots the agent, lets it produce a patch, applies the patch with git apply, and then runs the verifier. Result rows go back to the platform, which computes scores and exposes them at /scoring/weights. The validator periodically pulls that mapping and writes weights on-chain. The current production task set spans SWE-bench Verified (the 500-task human-validated subset of SWE-bench) and aider-polyglot (multi-language programming exercises). The recent commits show the team tightening operations: a cost-based tie-breaker between near-tied agents, short-circuiting evaluations on agents that fail early, and a stricter exclusion of cancelled or reference runs from the public leaderboard.
The on-chain side currently reads as winner-take-all in the strictest sense. Emission share sits at 0% of network emissions with the miner burn rate at 100% and 0 active miners reported by TaoSwap. The platform-mediated design means very few hotkeys hold the live agent slot at any time, so this picture is what the architecture produces by default rather than a sign of abandonment. Repo activity confirms this: development is clearly live, with the most recent commit on 2026-05-27. Market metrics are softer than they were when this subnet was first written up. Price is 0.01686 TAO with a market cap around 81,075 TAO, down about 38% on the month and roughly 61% over ninety days, while 7-day net flow is positive at about 624 TAO. Pool depth is around 36,540 root TAO and the root proportion is 0.166, both consistent with a relatively mature, organically traded subnet.
Two signals from social are worth flagging without overstating them. Public tweets in the last month claim a 73.6% score on SWE-bench Verified for a Ridges agent, and a 200 TAO single-trade alpha buy was called out on subnet-trade trackers. Neither is on-chain evidence the article verified directly, but both are consistent with a project that is still actively trying to ship rather than wind down.
- Sustained price decline: A 61% drawdown over ninety days is real. It reflects rotation out of the AI-coding-agent narrative and the lack of distributed reward at the protocol layer, even with active development.
- Winner-take-all and platform centralization: The Ridges platform is the source of truth for scoring, and the weights endpoint is a single dependency. That keeps quality high, but it means a platform outage or scoring bug has outsized impact compared to a more decentralized scoring scheme.
- Inference cost on miners: Required OpenRouter keys for both runtime and management mean a real-money cost to compete, on top of whatever the agent costs to run per task. This raises the barrier to entry for hobbyist miners.
- Centralized competition: Cognition, Cursor, GitHub, and Anthropic are well-capitalized and shipping fast on the same problem. The Ridges advantage has to be sustained open-source velocity, not a one-time benchmark win.
Into the next one.
