How the smartest engineers are building systems that build software — and why your next job title might be “Loop Architect.”
The Tweet That Changed Everything
On June 7, 2026, a dozen words broke the AI-coding internet. 26Peter Steinberger — the creator of OpenClaw, the open-source AI agent project that became the most-starred new repo in GitHub history — posted twelve words that ignited a firestorm across the software engineering landscape. The message was deceptively simple: “You shouldn’t be prompting coding agents anymore. You should be designing loops that prompt your agents.”25 Posted at 12:28 AM on June 8, 2026, the two-sentence post hit 6.5 million views on X. No diagram. No repo link. The entire AI-coding timeline spent the next week arguing about six words.
But it wasn’t just viral hype. Days before Steinberger’s tweet, 9Boris Cherny, head of Claude Code at Anthropic, had been saying the same thing publicly: “I don’t prompt Claude anymore. I have loops running that prompt Claude and figuring out what to do.” The old workflow — write prompt, read output, write next prompt — was being replaced by building small autonomous control systems that use the agents.
Two voices. Two sides of the industry. One converging conclusion: the era of prompt engineering is giving way to loop engineering.
So, What Exactly Is Loop Engineering?
4 “Loop engineering” is the name now attached to this shift, and it captures a real change in where the leverage lives: you stop being the person who prompts the agent and start being the person who designs the system that prompts it.
Put more precisely by Addy Osmani, Director of AI at Google Cloud: 17loop engineering is replacing yourself as the person who prompts the agent. You design the system that does it instead. A loop can be thought of as a recursive goal where you define a purpose and the AI iterates until complete.6 It’s the practice of designing AI systems that don’t just respond once — they act, observe the result, decide what to do next, and repeat until a goal is actually met.
This is a meaningful departure from the prior paradigm. 4Prompt engineering optimizes a single instruction you type by hand, one turn at a time. Loop engineering optimizes the autonomous system that decides what to prompt, when to prompt it, and whether the result is acceptable. Prompt engineering treats the agent as a tool you hold; loop engineering treats it as a long-running process with memory, scheduling, evaluation, and orchestration.
Where Does It Sit in the Stack?
17 Loop engineering’s sibling discipline is agent harness engineering — making the environment a single agent runs inside. Loop engineering sits one floor above the harness. The harness runs on a timer, it spawns little helpers, and it feeds itself. 4 It helps to see loop engineering as the third layer in a stack that has been building for a few years. Each layer wraps the one inside it, and each one moves the leverage point a little further away from the raw model call. 2 The progression: autocomplete (2023) → prompting AI to write code (2024) → running multiple agents in parallel (2025) → writing loops that do the prompting themselves (2026).
The Five Building Blocks of a Loop
19 Strip loop engineering down and you get roughly five building blocks, plus one place to remember things. Both Claude Code and Codex have all five now. The names differ here and there; the capability is the same.
1. Automations — The Heartbeat
4 Automations are what make a loop an actual loop and not just one run you did once. They are the heartbeat: a recurring trigger that surfaces work without you asking. Everything else in the loop reacts to what the automation finds. 28 Without a schedule, you have a one-off agent session. With one, you have discovery and triage on a cadence. This is the piece that turns “I should check CI every morning” into something that happens whether or not you open a terminal.
2. Git Worktrees — Isolation for Parallel Work
9 Two agents editing the same files at the same time is a merge disaster waiting to happen. Isolated git worktrees, or equivalent checkouts, give each agent its own working directory while sharing history. Both major coding-agent tools ship this, so sub-agents can be launched into fresh checkouts so parallel work does not collide. 13 A 2026 study on asynchronous coding agents found its gains — plus 26.7% on paper reproduction and plus 14.3% on library tasks — from isolating each agent in its own git worktree and verifying, not from simply adding more agents.
3. Skills — Persistent Knowledge
9 Every session, the agent starts cold. Conventions, build commands, review standards, and the lessons learned from hard incidents — all of it has to be re-derived unless you externalise it. Skills are how you pay down intent debt.
Addy Osmani has formalised this concept extensively. 23By encoding professional workflows, quality gates, and industry best practices directly into the operational logic of AI agents, agent-skills seeks to standardize how these autonomous systems interact with codebases — addressing a critical gap in the AI landscape, where the focus is shifting from simple code generation to the maintenance of high-quality, production-ready engineering standards.
4. Sub-Agents — A Maker-Checker System
4 The loop runs while you are not watching, so a verifier you actually trust is the only reason you can walk away. This is also what Claude Code’s `/goal` does under the hood: a fresh model decides whether the loop is done, not the one that did the work. The maker-checker split is applied to the stop condition itself. 9 Claude Code users have been observed running 10–15 parallel agents, using persistent memory files (CLAUDE.md / skills), sub-agents, and loops that scan GitHub, Slack, and Twitter, decide what to build, implement, test, and self-fix.
5. Connectors — Reaching into the World
24 The system reads a GitHub issue, triggers the agent, runs the test suite, catches the compiler error, feeds it back into the agent’s context window, and iterates entirely on its own until the tests pass. The system, not the human, becomes the driving motor.
6. Memory — The Spine of Everything
12 The sixth element is memory — a markdown file, or a Linear board, anything that lives outside the single conversation and holds what’s done and what is next. It sounds too simple to matter. But it’s the same trick every long-running agent depends on: the model forgets everything between runs, so the memory has to be on disk and not in the context. The agent forgets; the repo doesn’t.
A Loop in Action: Osmani’s Morning Workflow
Addy Osmani has described a concrete, real-world loop that elegantly demonstrates what all six parts look like assembled: 7a loop that each morning reads yesterday’s CI failures, the open issues, and the recent commits, writes findings into a markdown file, spawns an isolated worktree for each finding worth addressing, sends a worker agent to draft the fix and a reviewer agent to check it, opens the PR, updates the ticket, and puts anything it cannot handle in a triage inbox for the human. That is a full engineering workflow. The developer designed it once and did not prompt any of the individual steps.
Why Now? The Capability Inflection Point
The conversation about loops isn’t happening in a vacuum. There’s a hard technical reason this is the moment. 13Anthropic reports the task length a model can complete reliably is doubling roughly every four months, up from every seven, with its top model now handling jobs that take a human about twelve hours. More than 80% of the code Anthropic merges into its own codebase is now written by Claude.4 What surprised early adopters is that this is no longer a build-it-yourself effort. A year ago, a loop meant a pile of bash scripts you maintained forever and that only you understood. As of mid-2026, the pieces ship inside the products. 9 The tooling convergence is striking: both Claude Code and OpenAI Codex have landed on very similar primitives, so the “loop shape” is becoming somewhat tool-agnostic. This means engineers can design a loop architecture that is portable across tools — a powerful unlock for teams that don’t want to bet everything on a single vendor.
The Real Risks: What Industry Voices Are Warning About
Loop engineering is not a silver bullet. Some of the most respected voices in engineering are sounding important notes of caution.
⚠️ Comprehension Debt
13 Osmani calls it “comprehension debt”: the faster the loop ships code you did not write, the larger the distance between what the repository contains and what you understand. He pairs it with “cognitive surrender” — the pull to stop forming an opinion and accept whatever the loop returns.
As Addy Osmani himself puts it bluntly: 12when the loop runs itself, it is very tempting to stop having an opinion and just take whatever it gives back — what he calls “cognitive surrender.” Designing the loop is the cure when you do it with judgement, and the accelerant when you do it to avoid thinking — same action, opposite result.
⚠️ Weak Verification
13 Engineer Geoffrey Huntley documented the “Ralph Wiggum loop,” where an agent meant to emit a completion token only when finished emits it early, and the loop exits on a half-done job. Without a hard gate, loops fail quietly and keep spending.
⚠️ Security Vulnerabilities
13 A loop running unattended is an attack surface running unattended. A 2026 audit of 17,022 agent skills found 520 of them leaking credentials, with debug logging behind about 74% of the leaks. Skill descriptions double as a prompt-injection vector, since the agent reads them as instructions. A loop that auto-installs skills inherits every one of those holes without a human reading them first.
⚠️ Scale Isn’t the Answer
13 A 2025 survey of 306 practitioners across 26 domains (the Measuring Agents in Production study) found 68% of production agents run ten steps or fewer before a human steps in. The systems that work are small and supervised, not autonomous swarms.
Abhishek Murthy, Senior Principal ML and AI Architect at Schneider Electric, echoes this from the leadership perspective: 8“We have to be extremely careful in how much we trust and when we trust, making sure that there are some humans somewhere always validating what the AI systems produce.” His revised maxim? 8“Trust after verification, but verify continuously.”
Who Should (And Shouldn’t) Build Loops Right Now?
AlphaSignal’s analysis is among the most grounded in the field. 13Loop engineering pays off under four conditions: the task repeats, verification is automated, your token budget can absorb the waste, and the agent already has the tools a senior engineer would use. Miss one and the loop costs more than it returns.13 The teams that gain are those with repetitive, machine-checkable work and the budget to run it — continuous test triage, dependency bumps, lint-and-fix passes, issue-to-PR drafts on a codebase with strong test coverage. If a junior engineer could do the task from a checklist and a test suite would catch the mistakes, a loop fits. 13 The developers who should skip it are solo builders on consumer plans, anyone working on code with no automated verification, and teams whose real constraint is review capacity rather than typing speed.
The Bigger Picture: A New Kind of Engineering Role
24 Peter Steinberger is entirely correct: the highest leverage point for developers and enterprises today is moving up the abstraction layer to engineer loops. We are transitioning from writers of code to designers of autonomous machines that write code. 23 This shift could lead to a new era of “meta-engineering,” where the focus is on building the systems that govern AI behavior. Instead of manually performing every engineering task, developers may transition into roles where they define and refine the “skills” that AI agents use.
The engineering leadership community is feeling this at scale. 15After years of relative stability, the role of engineering leaders is changing rapidly to adapt to new economic and technological realities. Tightening budgets and the rise of generative AI are impacting how work is prioritized, the shape of engineering teams, and the way folks are rewarded for their work.
The Bottom Line
Loop engineering is not about replacing engineers. It’s about changing what engineers are responsible for. 4The goal is balance: set up loops for the recurring, verifiable work, and keep direct control for the parts where your judgment is the value. Build the loop like someone who intends to stay the engineer, not just the person who presses go.
The leverage point has moved. Prompting was the skill of 2024. 27The value is in moving judgment upstream, so the human designs the process and the model handles the recurring friction.
The question is no longer “how do I write a better prompt?” It is: “how do I design a better loop?”
🔗 Cited Sources
- Addy Osmani (Elevate/Substack) — Loop Engineering: https://addyo.substack.com/p/loop-engineering
- Addy Osmani (Personal Blog) — Loop Engineering: https://addyosmani.com/blog/loop-engineering/
- Addy Osmani (Personal Blog) — Agent Harness Engineering: https://addyosmani.com/blog/agent-harness-engineering/
- Addy Osmani (GitHub) — Agent Skills Project: https://github.com/addyosmani/agent-skills
- Cobus Greyling (Substack) — Loop Engineering: https://cobusgreyling.substack.com/p/loop-engineering
- Cobus Greyling (Medium) — Loop Engineering: https://cobusgreyling.medium.com/loop-engineering-62926dd6991c
- Cobus Greyling (GitHub) — Loop Engineering Patterns Reference: https://github.com/cobusgreyling/loop-engineering
- LushBinary — Loop Engineering: The Guide for AI Agents: https://lushbinary.com/blog/loop-engineering-ai-coding-agents-guide/
- MindStudio Blog — What Is Loop Engineering? https://www.mindstudio.ai/blog/what-is-loop-engineering-ai-coding-agents
- Pulumi Blog — Stop Prompting. Design the Loop: https://www.pulumi.com/blog/stop-prompting-design-the-loop/
- AlphaSignal (Substack) — Most Developers Do Not Need Agent Loops Yet: https://alphasignalai.substack.com/p/most-developers-do-not-need-agent
- AI Advances (Medium / Aftab) — You Shouldn’t Be Prompting AI Anymore: https://ai.gopubby.com/you-shouldnt-be-prompting-ai-anymore-you-should-be-designing-loops-5f26de35b84c
- Linas (Substack) — Loop Engineering: Design AI Loops That Ship While You Sleep: https://linas.substack.com/p/loop-engineering-complete-guide
- ExplainX.ai Blog — Loop Engineering 2026 Guide: https://explainx.ai/blog/loop-engineering-coding-agents-claude-code-guide-2026
- Filip Verloy (Medium) — From Prompt Engineering to Loop Engineering: https://medium.com/@filipv_74515/from-prompt-engineering-to-loop-engineering-why-the-agent-era-demands-a-new-security-paradigm-816385040e3d
- LoomStack — Loop Engineering Concept: https://loomstack.co/concepts/loop-engineering/
- Northeastern University News — Engineering Leaders Embracing AI: https://news.northeastern.edu/2025/10/21/engineering-leaders-embracing-ai-tech/
- LeadDev — Engineering Leadership Report 2025: https://leaddev.com/the-engineering-leadership-report-2025
- AIToolly — Agent-Skills by Addy Osmani: https://aitoolly.com/ai-news/article/2026-05-08-addy-osmani-launches-agent-skills-a-framework-for-production-grade-engineering-in-ai-coding-agents
- Digg — Peter Steinberger on Designing Agent Loops: https://digg.com/ai/7ifyvmb9