Rails, fundamentals, and AI to make things happen.

Conventions make parts of a system predictable. A name suggests where to look; a route points to a resource; a familiar pattern helps explain a new area. An agent can use these signals to navigate the code. When an application maintains that coherence, its intent becomes easier to find and connect to behavior.

Garry Tan, of Y Combinator, publicly praised Rails paired with Claude Code. In his post, he connects Rails' expressiveness with code generation by language models. That observation provides a starting point for exploring how a language and its ecosystem can support work with agents.

Ruby's expressiveness lets you represent ideas with little noise, and the talk presents studies of token efficiency. A task's cost also depends on the context loaded, searches, and attempts required. Assess the complete journey of finding the behavior, understanding the change, implementing it, and verifying that it solved the problem.

Knowing the framework helps with writing and locating code. Delivering a feature also requires understanding the product, respecting constraints, and verifying the expected behavior. The Agents on Rails studies distinguish framework knowledge from feature implementation. That difference helps you choose what to ask an agent and how to assess its delivery.

The study considers context cost, discoverability, isolation, predictability, and the reach of a change. These criteria help examine how each version organizes work for an agent. The scores describe this app and the tasks analyzed. Examples, files, and decisions let you investigate the results and form your own judgment.

A unified vocabulary connects the application's layers. Declaring responsibility for rules gives decisions a home. Naming orchestrations makes the steps of an operation visible. Organizing resources reduces exceptions in routes. You can follow these changes in the branches, read their reasoning, and compare the consequences in the code.

Structure can bring isolation while also creating new search paths, contracts, and coordination rules. The experiment shows this tradeoff. Increasing sophistication does not produce continuous improvement in Agent-Friendliness. The choice must consider the need a boundary serves and the effort it adds to understanding and changing the system.

Instructions record decisions, constraints, and paths that code alone does not explain. Clear names and responsibilities reduce the guidance needed to compensate for confusing structure. Use documentation to preserve important context and support discovery. The code should also communicate its concepts and make it clear where each behavior takes place.