How do you like to work with AI coding agents?
There seems to have two types of developers:
Human in the loop: Those who like to control the behavior of their agents as it works, looking at the context usage, reading reasoning blocks, and approving individual file edits.
Agent first: Those who prefer to review the output of agents, rather than individual actions, and run one or more sessions in parallel.
What type of developer are you when working with AI coding agents?
For context, when @Kilo Code launched their new @VS Code extension last week, with parallel agents, inline diff reviewer, and multi-model comparisons, some users actually wanted more control back.
While the team is working on multiple improvements, like better diffs before you approve and permission flow fixes, I'm curious about your point of view.
How do you like to work with AI coding agents?
Replies
I prefer a hands-on workflow because the projects I’m working on tend to be detailed and change as they develop. Staying involved in the steps lets me catch issues early, adjust direction on the fly, and make sure important context is incorporated before the work moves too far ahead. For me, that reduces the chance of agents completing a large amount of work that later has to be reworked because a detail, assumption, or workflow decision was missed.
I personally prefer Human in the loop since it helps control the quality output of my code. But AI still useful a lot in lots of basic and automation steps
I end up in a hybrid mode. For low-risk refactors, agent-first is fine as long as the run has a clear stop condition. For architecture, migrations, or anything touching auth/data, I want human-in-the-loop before the next attempt is allowed, not only after the final diff.
The useful UI for me is less "show every thought" and more: what changed, what evidence passed, what failed, what budget/context was consumed, and why the agent is allowed to keep going. Better diffs before approval solve part of that; explicit halt reasons and run receipts solve the rest.
slowly and meticulously, step by step one little snippet of code at a time.
Definitely human in the loop for me. I mainly just mess around with small game clones in Python or p5js so I actually need to read the logic to understand what is happening. Letting an agent just build everything in the background kind of defeats the purpose of trying to upskill and learn the code myself.
I am working with AI agents both in my day job and in my side project UnBullet (AI job matching platform). They are super useful and allow us to progress far faster and resolve problems than we ever imagined. We can also automated boring tasks like documention and stuff. So thats great. However, I strongly believe they are not yet ready to be trusted on their own. I have witnessed multiple occassions where a simple, strightforward task turns into a nightmare, really unexplicably. Also the code writing style, man so different from a human being. And the pace of writing code, lol who is gonna check or test all this stuff thoroughly? And most importantly, when something breaks (it will, it always does eventually), who are you gonna blame? The AI? The instructor? The management who took the decision to replace part of their human staff with agents? Difficult questions to answers and could get philosophical real quickly but the bottom line is: as of today May 24, 2026, I would not trust an AI agent to fully independently and autonomously work on comple projects end-to-end.
The choice often comes down to test coverage more than personal preference. Good tests let you go agent-first and use the test suite as your checkpoint. No tests means you’re forced into human-in-the-loop because you have no other reliable signal that nothing broke silently.
I'm human-in-the-loop, but only because I don't trust the review step yet. The day I can check an agent's result in one glance instead of reading every edit, I'd go agent-first. The bottleneck isn't the agent, it's how fast I can verify its work.