Why AI coding agents feel powerful at first, then become harder to control
3 comments
I wrote a longer piece exploring this, but I’m more interested in hearing how other engineers are approaching it in practice.
https://codeaholicguy.com/2026/01/31/ai-coding-agents-explai...
https://codeaholicguy.com/2026/01/31/ai-coding-agents-explai...
I have also encountered similar problems. Is it possible that the current AI coding is difficult to converge to the state we desire? Therefore, I think that currently, the AI tools are more suitable as a conversational tool when dealing with large projects. For example, humans can envision the framework and break down the implementation steps. Then, each small task can be given to the AI to solve, and then humans can integrate the code. This efficiency is lower than complete vibecoding, but it can alleviate the situation where the code becomes unmanageable.
I agree with this.
Right now, models are good at solving small, local problems, but much weaker at keeping large systems aligned over time. So having humans own the overall design, break work into small tasks, and integrate the results is a very pragmatic approach.
I see this less as a permanent limitation and more as a workflow gap. When AI is used purely as a conversational tool, humans end up doing all the convergence manually.
Concepts like rules, skills, scoped agents, and verification feel like early attempts to move some of that convergence into the system itself, not to replace human judgment, but to reduce how much needs to be constantly reapplied.
Right now, models are good at solving small, local problems, but much weaker at keeping large systems aligned over time. So having humans own the overall design, break work into small tasks, and integrate the results is a very pragmatic approach.
I see this less as a permanent limitation and more as a workflow gap. When AI is used purely as a conversational tool, humans end up doing all the convergence manually.
Concepts like rules, skills, scoped agents, and verification feel like early attempts to move some of that convergence into the system itself, not to replace human judgment, but to reduce how much needs to be constantly reapplied.
At small scale, prompting works. As the codebase grows, things start to drift: - constraints leak into prompts - commands get longer - behavior becomes inconsistent - “helpful” changes appear outside the intended scope
Lately, I’ve been mapping common agent concepts (rules, commands, skills, sub-agents, MCP, hooks) not as features, but as control layers in an execution loop.
Very roughly: - rules constrain decisions - commands trigger execution - skills encode repeatable methodology - sub-agents limit responsibility - MCP enables observation - hooks enforce guarantees
My hypothesis is that many frustrations come from mixing these layers.
For example, encoding constraints in commands, or trying to use hooks for reasoning.
I’m curious how others here are thinking about this: - Do you treat agents as conversations or as systems? - At what point does prompting stop scaling for you? - How do you control blast radius when agents touch larger codebases? - Have you found concepts like skills or sub-agents actually useful, or just extra complexity?