I’ve been spending a lot of time working with AI coding agents lately, and I keep running into the same pattern.
At the beginning, everything feels great.
You write a few prompts, code is generated, and velocity goes up immediately.
Then the project grows.
- Prompts get longer.
- The agent starts touching files you didn’t intend to change.
- Constraints slowly leak into prompts.
- The workflow becomes harder to reason about.
Nothing is “broken”, but everything feels fragile.
I don’t think this is mainly a model problem.
I think it’s a mental model problem.
Most of us still treat AI coding agents like conversational tools. In practice, they behave much more like execution systems. They read context, plan, execute, sometimes observe results, and repeat.
Concepts like rules, commands, skills, sub-agents, MCP, and hooks keep showing up in different tools. I don’t see them as features. I see them as attempts to control different parts of that execution loop:
- constraints
- execution triggers
- reusable methodology
- scope and responsibility
- observation and verification
- enforcement
A lot of pain seems to come from mixing these layers together. For example, encoding constraints inside prompts, or using automation to “reason”.
My current workaround looks similar to what many people describe:
- humans own the overall design
- break work into small, clear tasks
- let AI solve those locally
- integrate carefully
It’s slower than full vibecoding, but much more predictable.
I’m curious how others here are approaching this:
- At what point do prompts stop scaling for you?
- Do you structure your agent workflows, or keep things conversational?
- Have skills or scoped agents helped, or just added complexity?
I wrote a longer piece exploring this more deeply, but I’m more interested in hearing real experiences from other engineers.
Top comments (0)