What a coding agent does in a software engineering project
We use Claude Code and Cursor as coding agents, and they work inside the client's repository rather than in a chat window. Given a task, the agent reads the surrounding code first: the module it will change, the types it depends on and the tests that already cover it. It then plans the change, implements it, writes or updates tests, and runs the type checker, linter and test suite before proposing anything. The work it suits is wide and repetitive: wiring a new field through an API handler, its validation schema, the TypeScript types and the admin form; writing the migration for a schema change an engineer has already designed; adding tests to a module that never had any; or renaming a concept across hundreds of files without missing a call site. What these tasks share is that a person has already made the design decision. The agent applies it consistently across many files, and the compiler and tests confirm that every call site was updated.
