Great breakdown of the current AI coding landscape, Patrick. I've been deep in this space and your observation about "junior programmer-like collaborators" really resonates. What I've found is that the gap between tools isn't just features - it's how they handle context and maintain coherence across a real project.
I've been building an AI agent system using Claude Code for the past few months, and the human-in-the-loop dynamic you describe is spot on. But I'd add a nuance: the quality of that loop matters enormously. With some tools, you're constantly course-correcting. With others, you're more like a senior dev doing code review - which is a much more productive collaboration pattern.
The reliability issues you mention are real, but I've noticed they correlate heavily with how you structure the interaction. Giving AI full context about your codebase architecture, keeping focused scope, and treating it as a conversation rather than a command-line - these patterns dramatically reduce the "hallucination and unreliability" problems.
One thing missing from most comparisons is how these tools perform on sustained, multi-file projects versus one-off tasks. The demo experience and the daily-driver experience can be wildly different. I actually wrote up my findings after weeks of real-world testing with Claude Code specifically: https://thoughts.jock.pl/p/claude-code-review-real-testing-vs-zapier-make-2026
Would love to hear what your readers are finding works best for larger codebases - that seems to be where the tools really diverge in usefulness.
AI coding has improved massively since I wrote that Review. Much has changed.
Best solution: Claude Code, built on the same paradigm as Aider (terminal-based) is now the best-in-class coding assistant for large programming tasks, and it can do so much more thanks to Claude Opus 4.5 combined with use of skills and tool capabilities. AI coding assistant was helping on line-by-line in 2023, function-by-function in 2024, multi-file editing in 2025, and now can do whole projects.
Personally, I am using Claude Code and Roo Code in VS code for my AI dev work. It's a good combo to have an in-IDE solution and a terminal one as well, based on the task at hand.
Great breakdown of the current AI coding landscape, Patrick. I've been deep in this space and your observation about "junior programmer-like collaborators" really resonates. What I've found is that the gap between tools isn't just features - it's how they handle context and maintain coherence across a real project.
I've been building an AI agent system using Claude Code for the past few months, and the human-in-the-loop dynamic you describe is spot on. But I'd add a nuance: the quality of that loop matters enormously. With some tools, you're constantly course-correcting. With others, you're more like a senior dev doing code review - which is a much more productive collaboration pattern.
The reliability issues you mention are real, but I've noticed they correlate heavily with how you structure the interaction. Giving AI full context about your codebase architecture, keeping focused scope, and treating it as a conversation rather than a command-line - these patterns dramatically reduce the "hallucination and unreliability" problems.
One thing missing from most comparisons is how these tools perform on sustained, multi-file projects versus one-off tasks. The demo experience and the daily-driver experience can be wildly different. I actually wrote up my findings after weeks of real-world testing with Claude Code specifically: https://thoughts.jock.pl/p/claude-code-review-real-testing-vs-zapier-make-2026
Would love to hear what your readers are finding works best for larger codebases - that seems to be where the tools really diverge in usefulness.
AI coding has improved massively since I wrote that Review. Much has changed.
Best solution: Claude Code, built on the same paradigm as Aider (terminal-based) is now the best-in-class coding assistant for large programming tasks, and it can do so much more thanks to Claude Opus 4.5 combined with use of skills and tool capabilities. AI coding assistant was helping on line-by-line in 2023, function-by-function in 2024, multi-file editing in 2025, and now can do whole projects.
Personally, I am using Claude Code and Roo Code in VS code for my AI dev work. It's a good combo to have an in-IDE solution and a terminal one as well, based on the task at hand.