Coding Agents Write Swift Well, But They Can't Finish It

𝗖𝗼𝗱𝗶𝗻𝗴 𝗔𝗴𝗲𝗻𝘁𝘀 𝗪𝗿𝗶𝘁𝗲 𝗦𝘄𝗶𝗳𝘁 𝗪𝗲𝗹𝗹, 𝗕𝘂𝘁 𝗧𝗵𝗲𝘆 𝗖𝗮𝗻'𝘁 𝗙𝗶𝗻𝗶𝘀𝗵 𝗜𝘁

I spent months testing AI coding agents on real Swift and Xcode tasks. I did not use simple demos. I gave them real tasks with build targets and tests.

The first draft is usually fine. A capable model writes code that looks correct. The types line up. The views make sense. If writing Swift was the only problem, these tools would be finished.

The real issue happens after the first draft. It is the gap between code that looks finished and code that is actually right.

Modern tools like Claude Code or Codex handle compile errors well. They do not hand you code that fails to build. The real failures are the ones the compiler cannot see.

These failures waste my time:

• Intent mismatch: The code compiles and tests pass. However, the behavior is wrong. The agent follows the green light of a passing test instead of your actual goal. It has no way to check intent.

• Concurrency races: The code builds clean but has data races. These only appear during runtime. The agent sees a green build and moves on. Fixing these usually requires a redesign, which the agent fails to do.

• Regression loops: The agent fixes one bug but breaks another. While fixing the second bug, it undoes the first fix. It enters a loop where it oscillates between two broken states. It lacks a memory of what failed before.

• Poor architecture: The agent writes code that fights the framework. It ignores your app structure. It works for a throwaway script but fails in a real codebase.

These are not language problems. The model knows Swift. It misses what the compiler cannot see. It does not know if the result matches your intent or if the structure is professional.

This shifts the cost from tokens to attention. An agent that requires you to watch it every few turns has not saved you work. It converted writing into supervising.

I am improving my results by changing the loop around the model. I focus on what the agent checks and remembers.

Does this match your experience? When you run agents on Apple platforms, where does it break? Is it intent mismatches, runtime races, or something else?

Source: https://dev.to/jbrackin/coding-agents-are-good-at-writing-swift-theyre-bad-at-finishing-it-md3

Optional learning community: https://t.me/GyaanSetuAi