Rails gives coding agents the same thing it has always given developers: clear conventions, expressive code, and a complete framework for turning ideas into production software.
Rails gives agents a map. Standard names, folders, commands, and patterns help generated changes land closer to idiomatic Rails with less prompting.
Less code means more context. Ruby and Rails express product ideas with fewer tokens, helping agents make smaller edits and move faster from request to working feature.
The patterns are everywhere. Decades of public Rails code give models strong signals for controllers, models, views, tests, jobs, migrations, and the glue between them.
One person can ship more. Rails provides the full product stack. Add an AI coding agent, and a solo builder can take on broader work without stitching everything together from scratch.
One benchmark, from small self-contained tasks to work that spans multiple steps.
Compare models by accuracy, speed, token efficiency, and cost, on task types that matter most to you.
Select one or more task types. All are selected by default.
The strongest models rise toward the top-right.
We compare models on small tasks in Writebook, feature tickets in Fizzy, and multi-step work in new and existing Rails apps. Every run uses the same lemans harness and miniswen agent, with a single bash tool.
Hidden tests score each run from 0.0 to 1.0. A failed required check earns zero; once all required checks pass, optional checks add credit based on their importance. We repeat each task and average its runs first, so every task has equal weight. Accuracy is the average task score as a percentage. Tokens, cost, and speed are averaged across tasks too.
Each row shows a model at the listed reasoning effort. Colors compare configurations for each metric within your selected task types: green marks the top 20%, yellow the next 30%, and red the rest. In model details, scores compare configurations on each individual task.
Select a result to see task scores and run counts, or explore the open-source Rails AI evaluation suite.
Convention over configuration set the path for 20+ years of great training data for AI to use today. Not only does this mean agents do great with Rails, but also that squishy humans can quickly and confidently review the output without a jungle of distracting boilerplate.
Rails' convention over configuration explains why I got very good LLM output quality in the early days of LLM-assisted coding while many developers still thought it was unusable. Most Rails codebases look the same and on average is high quality.
One of the killer features of using Rails in this AI era is "Convention over configuration". If you keep your project close to the Rails defaults, the AI knows so much already about your project… It makes things 100x faster.