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
Human in the loop, always. I've been building an AI feature for Fluxerv where the model generates live UI components — streaming HTML into the editor in real time. If I let it run fully autonomous even for a few seconds without checkpoints, the output drifts fast. Review-as-you-go is slower but the results are coherent.
I have been using Claude Code, although sometimes I get stuck in circles when the coding is broken and often spend hours going back and forth. Any suggestions on how to fix this?
Human-in-the-loop, firmly. Building Composa solo — I let the agent run, but I want to see every diff before it touches anything structural. The moment I stopped reading reasoning blocks, I started shipping bugs I didn’t understand.
Agent-first, but with one rule: I read every file it touches before it ships. Built 5 Shopify apps with Claude Code — Remix, Prisma, Polaris. The biggest unlock wasn't speed, it was being able to hold the entire app's context in one session. I'm a solo founder. Without it I'd need 3x the time or a co-founder.
Best setup for me is hybrid: parallel agents + quick review of diffs, not micromanaging every action.
I’m mostly hybrid. Agent-first is great for bounded work like refactors, boilerplate, test expansion, or exploring a few approaches in parallel.
For architecture, migrations, auth, billing, or anything with data risk, I still want human-in-the-loop with a reviewable diff before approval. The trust comes less from “the agent is smart” and more from predictable checkpoints.
Somewhere in between depending on what we're building.
For anything touching production at Gaper or Altstack the rule is to keep a human in the loop, no question. I want to see the reasoning, approve the file edits, and know exactly what changed before it goes anywhere near real systems.
For prototyping or internal tooling where the stakes are lower, agent first. Let it run, review the output, iterate fast.
I’ve moved toward agent-first workflows, but the real bottleneck becomes verification. Once you run multiple sessions in parallel, reviewing every action defeats much of the speed advantage, while reviewing only the final output can hide bad assumptions or incomplete work.
The missing middle is letting agents work independently while enforcing scope, verification, retries, and a clear evidence trail before accepting the result.
as much as going completely agent-first is wanted, its still a dream, agents are not ready and need a lot more governance and controls like what martinloop provides, open source.