Technical Context
I’ve long waited for someone to say the obvious out loud: manually prompting coding agents is starting to feel like yesterday’s news. Peter Steinberger wrote exactly about that, drawing on his experience with OpenClaw: you don’t need to polish the perfect prompt, you need to build a loop where the agent receives a task, takes a step, gets checked, and moves to the next iteration.
And this already looks like proper AI architecture, not shamanistic prompt engineering. In practical terms, his OpenClaw acts as a supervisor over multiple Codex agents, shifting focus from 'what else to add to the prompt' to 'what sensors, checks, and launch rules I’ve baked into the system'.
I’d describe the idea like this: the agent is no longer trusted on its word. It’s run through compilation, tests, linters, runtime errors, screenshots, API responses, and other signals that feed it back into the loop. The human doesn’t disappear but stops being the operator of every command and becomes the architect of that loop.
What especially caught my eye was the supervision aspect. In the post’s discussion, an author-made config for Paperclip surfaced, featuring two levels of control: a conditional CTO supervisor coordinates development and measures efficiency, while engineering-ops later analyzes logs and metrics and suggests how to change agent skills or the Agents.md file. That’s starting to look like a mature system, not a weekend toy.
What This Means for Business and Automation
First: the winners are teams that need repeatability, not magic. If I’m integrating AI into development, I care less about ‘wow, the agent wrote a feature on its own’ and more about whether the system reliably catches errors before they hit production.
Second: the value of the surrounding framework skyrockets. The model itself is no longer the center of the universe; the center is the loop, checks, task routing, and supervisor roles. Those who still measure quality by the number of prompts in Notion are losing out.
And third: parallel work by multiple agents becomes more realistic. But only if someone has designed the loop properly, otherwise you get not AI automation but an orchestra of hallucinations.
That’s exactly the kind of thing I build in my work: not just plugging in a model, but turning it into a managed process. If your development, support, or internal ops team has hit the ceiling of manual routine, feel free to break it down at Nahornyi AI Lab and assemble an AI solution development tailored to your real setup, without any faith in magic prompts.