A software engineer reports that AI models seem to be getting worse at following instructions in recent updates, often making unasked changes and ignoring contracts, leading to increased manual work.
Hello, I am a software engineer who uses a workflow where I would develop an extensive white doc and plans for building new features, hand this to AI, then review and correct from there. I have found over the past year moving to an AI-heavy workflow, that it feels like frontier models are getting worse at following instructions. I have an AI contract that I continuously update, but now I am dealing with agents not reading them at all. I am constantly finding places where AIs will change things totally unrelated to my white docs and plans, usually UI changes, things I never asked for, even with a contract that forbids them from making any code changes without explicit permission. These models are supposed to be improving, but I find that they are making more unasked for changes with each release, they are getting worse at following instructions. Has anyone else noticed this? I will give AI a white doc that tells it exactly what to code, and it will still do things outside of the white doc when implementing it, I did not have this happen as often with older models, I'm talking the models prior to Chatgpt 5.1. It seriously feels like sometimes it's intentional, like if you make a bug every turn, it may be a way to keep users engaged. How are other engineers navigating these issues? I use hooks and contracts, the AIs seem to ignore them, and confidently act as though they did review them, until I point out how they break the rules. I have tried both long and concise contracts, the problems seem to be getting worse with newer releases, not better, even with documents that tell the AI exactly what to code, it will code something that's not asked for, leading to me manually fixing it. It's almost like it just isn't possible for the AI to do anything anymore without either added in a unasked for change, even when given a plan describing exactly what to code, it seriously feels like the models I am using are not the ones they run in benchmark tests. Does anyone else feel this way? How do you deal with it? Both short and long white docs and contracts , even scripts get ignored, so every prompt results in something left unfinished or a bug added , just so you spend more time on the agent. That's how it feels, older models followed my instructions better, I find myself using older models as a result more often than the latest ones. AIs seem to go as far as ignoring contracts, not running tests and scripts to just add one unasked for change every turn, it seriously feels intentional. It doesn't feel like I'm working with anything like the benchmark tests say I am working with. Even minor fixes and changes seem to result in the AIs doing something along the way just so there's one more issue for me to fix. I am considering running some tests to see if this is a programmed behavior, how far would the AI model go to disregard the user's prompt just to make sure that the user is never able to leave the session with a totally finished job? Let me know if it's just in my head, it just feels like the models are getting worse and not better, older models seem to obey and read contracts better than newer ones, especially after Chatgpt 5.1 and Claude Opus 4.8. If you experience the same issues, how do you deal with them? I've tried both short and long white docs, contracts, tests, scripts, hooks, etc. I use Codex and Claude Code, working on my own custom Hermes-based agent so I can have full access controls for local models.
A long-time AI user reports that models across ChatGPT, Grok, Claude, DeepSeek and coding agents like OpenCode have noticeably worsened at task understanding and instruction-following over the past year, prompting questions about post-training, alignment and inference changes. The post asks the community whether others are seeing the same cross-provider regression.
A developer reflects on how AI coding agents can rapidly build and modify software, but developers often lose understanding of the codebase architecture and decisions, creating a new engineering challenge.
The article critiques the trend of giving AI agents more autonomy with less human oversight, arguing that it ignores hard-won software engineering practices like code review and staged rollouts, leading to silent failures and unexpected costs.
AI has made writing code cheaper but shifted the hard work to context-setting, review, and cleanup, requiring more skilled oversight. The article argues that teams often treat AI-generated code as finished work when it is actually a fast first draft.
A practitioner argues that AI agent builders are rediscovering problems that basic software engineering—input validation, data modeling, process isolation, access control—already solved, and that vibe coding skips these cross-cutting concerns. The post calls for applying established design and architecture principles to agent workflows.