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
For me it's a little bit of both... When using antigravity, I usually ask my AI agent to create a succinct implementation plan for whatever it is I am trying to do, then I review it, make comments within it, and if I am satisfied, I let the agent do its thing.
Definitely human in the loop - and it's not because I don't trust LLMs. I came to the conclusion that LLMs can do a great job without me having to supervise them, but there are two big reasons:
I still want to know what I am going to commit and send (especially when working as an outsourcer)
I want to be in the loop in terms of the knowledge of the codebase (i.e. I need to understand what are the sources of some potential issues/unexpected behaviour etc.)
I might need to make some small adjustments to match the existing codebase
All that being said, I am sometimes going into the agent-first mode when working on pet projects - mostly for challenging the process and seeing how far I can go :)
I'd probably consider myself a "High-level Human-in-the-Loop." I don't approve every single edit, but I prefer to keep an eye on the overall direction.
Since I work a lot with a fast-evolving runtime like Bun, AI models can sometimes be a bit behind the latest updates, so some human oversight is still quite necessary for me. Plus, I notice agents often tend to over-complicate things - like adding too many CSS classes where clean cascading would do the job.
For me, a hybrid approach feels like the right balance: letting the agent do the heavy lifting, but stepping in manually when it's just faster (and cheaper on tokens) than writing yet another prompt.
I’m somewhere in the middle agent-first for scaffolding and repetitive work, human-in-the-loop for architecture and risky edits. The sweet spot is when the agent moves fast but still gives me clean checkpoints to review instant of micromanaging every token.
I'm a bit frustrating about the new version of Kilo Code. The agentic workflow make me completly blind of what's going on. Tokens running and at the end of the task I'm not sure that agent was handled the given mission in the right way. Still working with the previous "old school" version that I can interrupt / correct on each step that I desire.
Tabstack by Mozilla
appreciate the feedback!
The new @VS Code extension launched this month pushed hard on capability - parallel agents, inline diff reviewer, multi-model comparisons. Some of that came at the cost of visibility and control. The team has worked on rebalancing it with multiple improvements, like expandable reasoning blocks, better diffs before you approve, and permission flow fixes.
This is a good time to check back. The experience is meaningfully different now. cc @realolearycrew
kilo.ai/install
people said the same thing about AI writing copy and now half the internet runs on it without a second look. trust builds with consistency not permission screens
Cosmic
Both modes matter at different stages, HITL when exploring a new domain, agent-first once the patterns and guardrails are trusted. The teams that get stuck are the ones who stay in review-every-line mode indefinitely. Context richness usually determines when you can safely let go.
Coming into the convo a bit late, but I'm also one of the "it's got to have both parts". The Agent can type really fast, which is great...but having the human present and running the show is important - I'd say even vital - to getting a quality product.
I look at the agents like a hammer. You can build a house without a hammer, but it's going to take a lot longer. The agents - being a "tool" in the toolbox - make the project faster and (at least for me) better quality without the human-induced errors likely when we're tired, distracted, hungry, etc.