跪拜 Guibai
← All articles
Artificial Intelligence · Android · GitHub

Loop Engineering Is the Next Rung After Prompting Coding Agents

By 朱涛的自习室 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

The abstraction leap from prompting agents to designing autonomous loops compresses a workflow change that previously took years into months. Teams that adopt Loop Engineering shift their bottleneck from prompt-crafting to verification design, and for any shop shipping mobile or web UIs, the missing piece is an automated real-device testing loop that doesn't require a human to tap through screens.

Summary

A six-word post from OpenClaw's Peter Steinberger — "design loops that prompt your agents" — and a follow-up framework from Google's Addy Osmani have formalized a shift in AI programming. The core idea: stop typing prompts turn-by-turn and instead build recursive systems that prompt, verify, and re-prompt agents autonomously. Claude Code's lead Boris Cherny confirmed the pattern is already running in production, with Claude Code now 100% self-maintained and responsible for roughly 4% of public GitHub commits.

A working Loop combines automations for scheduled discovery, isolated worktrees for parallel tasks, skills files for persistent project knowledge, connectors to external services, and separated sub-agents so the coder doesn't grade its own work. The bottleneck for mobile and web teams is verification: lint and unit tests can gate a merge, but checking whether a UI actually works on a real device still requires a human to tap through screens.

Munk AI, a new open-source tool, attempts to close that gap by giving coding agents a verification sub-loop that controls real Android and iOS devices. It reads requirements and code diffs, executes user paths on-device, and returns structured failure evidence — screenshots and step logs — so the coding agent can fix issues and re-run the check until the product actually works.

Takeaways
Loop Engineering means designing systems that autonomously discover work, dispatch agents, verify output, and iterate — instead of typing prompts one turn at a time.
Claude Code is already 100% self-maintained, and roughly 4% of public GitHub commits come from it, proving the loop pattern works at scale.
A functional Loop requires five components: automations for scheduled triggers, isolated worktrees for parallel tasks, skills files for project knowledge, connectors to external services, and separated sub-agents so the implementer doesn't also judge its own output.
State stored on disk — not chat history — is what lets a Loop resume across sessions without losing context.
Most coding Loops stop at "code is mergeable" because verifying a UI on a real device still requires a human; lint and unit tests can't confirm a button is tappable or a flow is correct.
Letting a coding agent write its own Playwright tests creates a self-grading problem that undermines trust in the Loop's results.
Munk AI provides a verification sub-loop for mobile: it controls real Android and iOS devices, executes user paths, and returns structured failure evidence so the coding agent can fix issues and re-run checks automatically.
Conclusions

The jump from prompting to loop design is a leverage shift, not a labor-saving one — the work doesn't get easier, but the point of highest impact moves from writing good prompts to designing good stopping conditions and verification gates.

Verification is the hardest unsolved piece of autonomous coding Loops, and the three obvious paths — manual testing, self-written UI scripts, and cloud vision models — each fail on cost, trustworthiness, or scalability.

Mobile and web teams hit the verification wall first because their completion criteria are visual and interactive, not terminal-based; a green CI pipeline says nothing about whether a confirmation dialog was actually tappable.

Separating the maker agent from the checker agent is not optional for unattended Loops — a single model grading its own output is structurally biased, and the Loop's credibility collapses without that separation.

Concepts & terms
Loop Engineering
A paradigm where developers design autonomous, recursive systems that prompt, verify, and re-prompt coding agents until a goal is met, rather than manually prompting agents turn-by-turn.
Harness Engineering
The practice of building environments, rules, and feedback mechanisms that let a single AI coding agent run safely and reliably — the 'runway' on which a Loop operates.
Verification Sub-Loop
An automated branch alongside a coding Loop that independently checks whether the product actually works on a real device or browser, using separate agents for execution and judgment, and returning structured failure evidence.
From the discussion

The core tension is cost: autonomous loops burn tokens, and someone has to foot the bill. The reply frames this as the new normal — developers are already effectively paying to work. Other comments stay on the surface: one calls the article substantial, another jokes about delaying learning until a fortune is made, and a third quips that AI's pace makes learning optional. A project link rounds out the thread.

Token cost is the immediate practical objection to autonomous agent loops.
The counterpoint is that developers already absorb work-related costs, so token spend is just another form of 'paying to work.'
AI's rapid advancement creates a perverse incentive: if you can't keep up, there's no point starting.
Featured comments
老年白兔码农 1 likes

Who pays for the tokens?

朱涛的自习室  · 2 likes

Nowadays, isn't everyone just 'paying to work'? [sinister grin]

magicwing

/goal Let's set a small target first: earn a hundred million [grin]. I'll learn it after that's done. A bird in the hand is worth two in the bush.

日是故乡明

In the AI era there's a famous saying: as long as you don't learn fast enough, you don't need to learn anything at all!

See top comments, translated →
Source: juejin.cn ↗ Google Translate ↗ Backup ↗