PR for a Ticket

The City upon a hill for AI coding is that a well specified ticket should get an automated pull request implementing it. The act of coding itself will go from carpentry to CNC, where you still must understand the properties of the material you work with, but shaping that material is a easily repeatable task, so you just focus on the goal.

Many of the day to day tickets a software engineer runs into are already at the point where a well instrumented AI Agent can handle the changes. Updating a title, creating a well specified screen, and even implementing a simple CRUD endpoint are almost fully solved. However, many tasks, like creating a database migration, managing security concerns, and other tricky tasks are beyond current agents.

The problem is there’s a discontinuity on AI solutions solving the former and the latter. You have tools like v0, which are the specification first agent, designed to start with a greenfield project and do very simple things. Break out of the rails, and thing become very difficult. Then other site of the fence is Cursor, Windsurf, Cline, which are amazing tools for the developer, but don’t even attempt to solve for full automation.

This is why Claude Code and Codex excite me. They’re able to be used in a continous fashion. If you instrument a codebase to be well used by them (Docs, MCP, rules), then it becomes relatively trivial to use them headlessly.

Much like the models before them, the coding agent becomes another layer of abstraction. Instead of opening your codebase and having one thread of “make a plan…” then another of “do the plan…” and lastly “test and merge the plan…”, you can have a workflow tool do those prompts automatically.

I think Cusrsor and friends are dead ends unless they can solve the last mile to get to full automation. Not every task will be automatable, but many will, and it should be a continuous system that handles local and remote development.