There is more than one way to work with AI agents
How you interact with these amazing new tools is up to you. We have collated some of the patterns of interaction that we have seen and that we use. We hope they are helpful. Please contribute if you have a different way.
The patterns
Tab Completion
In the early days of AI coding, this was it and it is where the agentic engineering journey started for many. Tends to be not used much as there are newer patterns that give a better combination of efficiency with control and safety.
Full Vibing with Agent
If you don't understand code at all or for whatever reason, you don't want to put hands on code for this project, you can use your coding agent to do everything and just not touch the code at all.
Full Vibing in a Walled Garden
If you don't understand code at all or for whatever reason, you don't want to put hands on code for this project, you can use helpful tools to do everything for you and hide the code and the agent from you. You just prompt and eventually you can get a system.
Plan, Execute, Check
A robust and professional approach. Good balance of speed and control. The current meta.
I Code, You Check
A High touch, high skill, professional approach for the most critical code. Complete control. You get to do the interesting bits. Not very fast.
More patterns to come
The rubric is fixed and the pages are generated, so a new pattern is a new markdown file — same card, same ratings, same page shape.
Side by side
The same judgements for every pattern, in the pattern's own terms. Brainrot and token use read the other way round from the rest — low is better.
| Pattern | Speed | Control | Quality | Safety | Brainrot | Token use |
|---|---|---|---|---|---|---|
| Tab Completion | Low | High | High | High | Low | Low |
| Full Vibing with Agent | High | Low | Low | Low to Medium | High | Highest |
| Full Vibing in a Walled Garden | Highest | Lowest | Low to Medium | Low to Medium | Highest | High |
| Plan, Execute, Check | Medium to High | Medium to High | Medium to High | High | Medium | Medium to High |
| I Code, You Check | Medium | Highest | Highest | Highest | Negative | Medium |
The six axes
Further out is better. Every rating below is a word written in the pattern's own markdown file, not a number typed into a spreadsheet.
Speed
How much wall-clock time the pattern actually saves — not how fast the tokens arrive.
Control
How small the increments are, and how easily you can steer, reject, or supply context.
Quality
What the code looks like a month later, once the session that produced it is forgotten.
Safety
The blast radius when it goes wrong, and how much of the damage you'd notice.
Brainrot
How much of the thinking leaves your head for good. Low is what you want.
Token use
What the pattern costs to run, per unit of work that survives review.