TL;DR
A developer who has used Copilot, Cody, Cursor, Claude Code, Codex and Gemini for roughly 18 months argues that engineers should adopt a single model and train it as a pair-programming partner. Coding assistants have improved in code quality, context grounding and sustained problem solving, but they still struggle with UI generation and require human oversight for product-quality software.
What happened
Over the past year and a half the author has incorporated several coding assistants—including Copilot, Cody, Cursor, Claude Code, Codex and Gemini—into day-to-day software work. They report three areas of notable improvement: models produce higher-quality code within supported languages, outputs are more faithfully grounded in the target codebase, and surrounding harnesses enable longer, coherent assistant sessions. The author finds assistants excel at routine business logic and bug hunting (citing Opus 4.5 and GPT 5.2 as strong debuggers) but are inconsistent on specialized one-off tasks or high-quality frontend UI code. They recommend treating a chosen model as a pair-programming buddy, investing time to teach and refine it, using assistants to handle toil, and creating external sandboxes and workflows to retain control. The author also flags risks: degraded long-term product quality when code is primarily agent-generated and erosion of an engineer’s internal mental model unless mitigated by tooling and training.
Why it matters
- Coding assistants can accelerate routine development and debugging, shifting engineers’ day-to-day work toward higher-level tasks and oversight.
- Relying heavily on agents without strong review skills and tooling may shorten the effective product-quality lifecycle of a codebase.
- Current assistants struggle with producing polished, idiomatic UI code, so front-end work still needs substantial human guidance.
- Teams will need new practices and tooling to preserve engineers’ mental models and to teach assistants project-specific norms.
Key facts
- The author has used Copilot, Cody, Cursor for about 18 months and began using Claude Code, Codex and Gemini after their recent releases.
- Three reported improvements in the last year: better code quality in supported languages, stronger grounding in the active codebase, and harnesses that enable sustained, coherent sessions.
- Assistants are effective at routine business logic and are often faster and higher quality than an average programmer under the author’s workflow.
- Models noted as strong bug hunters include Opus 4.5 and GPT 5.2, though they are not infallible.
- Front-end and UI frameworks remain a weakness: the author reports poor results with Tailwind, Ink, and Textual, and mixed results with Ratatui.
- Prompting for 'idiomatic' solutions and emphasizing intent in tests can steer model outputs toward better practices.
- The author uses an external sandbox (sandbox-exec) and disables built-in assistant sandboxes in some workflows to avoid coordination issues.
What to watch next
- Improvements in model training and harness design that target UI code generation and heavy abstraction frameworks (not confirmed in the source).
- Emerging tooling and training programs to help engineers maintain accurate mental models when assistants write much of the code (not confirmed in the source).
- Research or audits measuring code quality and long-term product outcomes for codebases generated primarily by coding assistants (not confirmed in the source).
Quick glossary
- Coding assistant: A system built around a language model that helps developers write, review, and debug code by generating or suggesting code and explanations.
- Agent: An autonomous component that performs specific tasks; multiple agents can be orchestrated together to form a coding assistant.
- Pair programming: A collaborative development practice where two programmers work together at one workstation, with one writing code and the other reviewing or guiding.
- Sandbox: An isolated execution environment used to run or test code safely without affecting production systems.
- Mental model: An individual’s internal understanding of how a system works, used to reason about behavior and make design or debugging decisions.
Reader FAQ
Are coding assistants ready to replace software engineers?
Not confirmed in the source.
Can these tools produce production-quality software on their own?
The source says code quality from assistants is necessary but not sufficient for product quality; human oversight and skills are still required.
Which areas do assistants struggle with?
According to the author, generating high-quality frontend and UI code (Tailwind, Ink, Textual) is a notable weakness; some UI toolkits perform better than others.
How should engineers start adopting coding assistants?
Begin by using assistants for repetitive toil—stack traces, summarizing docs, and navigating poor code—and invest time teaching and scaffolding the chosen model.
Programming, Evolved: Lessons & Observations Drafted December 2025 (v1), revised January 2026 (v2) TL;DR: If you are a software engineer, regardless of level, pick a model and shape it into…
Sources
- Programming, Evolved: Lessons and Observations
- AI Coding Assistants in 2026: How They're Transforming …
- AI Coding Assistants Are Reshaping Engineering
- Vibe Coding, AI Assistants and the Evolution of Software …
Related posts
- Gemini Personal Intelligence Offers a Preview of Google-Powered Siri Features
- Fine-tune an AI to write buggy code, and it begins fantasizing about enslaving humans
- NotebookLM is strong now – 5 features that could make it unstoppable