Boris Cherny, the head of Anthropic’s Claude Code, said in June that he has not manually written code in more than a year. Instead, he designs “loops”—recurring, self-checking workflows that let AI agents write code, run tests, and deliver reports while he’s away from the keyboard. The approach, now gaining traction as “loop engineering,” promises to shift developer work from directing one-off prompts to orchestrating autonomous, verifiable jobs. For Windows developers and IT administrators, the technique is not science fiction: it runs on existing toolchains like GitHub Actions, Azure DevOps, and PowerShell.
From Prompts to Persistent Agents
In a typical coding session, an engineer asks an AI for a change, reviews the diff, and iterates. Loop engineering flips this pattern. Cherny defines a task with clear success criteria, supplies the necessary tools and tests, and then lets an agent repeat until an independent verification pass. “My job is designing loops,” he told BigGo Finance. Those loops handle chores such as monitoring pull requests, automatically fixing continuous integration (CI) failures, repairing flaky test cases, and scraping social media for customer feedback every 30 minutes. He says he manages hundreds of agents daily, with thousands more running overnight.
The shift is enabled by Anthropic’s latest model, Claude Fable 5. Released earlier this summer, Fable 5 can sustain autonomous work for days rather than minutes, writes and executes its own tests after completing a task, and accurately interprets dense charts and diagrams. In Cherny’s workflow, two commands embody the loop philosophy. /goal sets a specific objective—for example, “ensure all unit tests in /tests/ pass and code style checks are clean.” After each step, a separate “scorer” model evaluates whether the goal is met. /loop, by contrast, runs a task on a fixed schedule until manually stopped—think “check error logs every 30 minutes and summarize critical entries.”
Crucially, the scorer is not the same model that performs the work. Cherny calls this division the “heart” of the system. Google engineering leader Addy Osmani, in a July 15 blog post, formalized loop engineering as the fourth generation of AI interaction, succeeding prompt engineering, context engineering, and harness engineering. Osmani argues that the core value is not the loop itself, but the ability to verify and explain what the system did.
What This Means for You
For individual Windows developers, loop engineering automates the most tedious parts of the job. Instead of manually chasing a CI flake or re-running a test suite every time a dependency shifts, you can offload those to a loop that reports back when a fix is ready or a new failure pattern emerges. It does not replace creative problem-solving, but it reclaims hours lost to repetitive checks.
For IT administrators and DevOps teams, loops can monitor system logs, triage support tickets, or draft reply templates—all within a controlled sandbox. The same orchestration tools you already use apply: a GitHub Actions workflow can invoke a loop on a schedule, an Azure DevOps pipeline can run a validation agent, and a PowerShell script triggered by Task Scheduler can serve as a lightweight harness. There is no new Windows automation framework to learn; the change is conceptual, not technical.
The risk, however, is real. An agent with broad permissions could consume paid model credits, weaken test assertions to make checks pass, or alter code unrelated to its task. Security-minded teams should never give a loop unrestricted repository, cloud, or production access. Scoped credentials, a finite budget, and human approval gates for merges or deployments remain essential.
How We Got Here
The road from simple autocomplete to autonomous agents has been short. Early AI coding assistants offered inline suggestions; then came chat-based copilots that could explain or refactor a block of code. By late 2025, tools like Claude Code could handle multi-file edits in a single prompt. Inside Anthropic, the culture shifted so far that even SQL queries are generated by models, according to reports.
Cherny’s personal timeline illustrates the acceleration. Fortune reported in June that he had not written code by hand for roughly eight months. BigGo Finance’s July 20 account extended that to more than a year. The discrepancy may reflect when he stopped certain types of manual coding, but the trend is clear. Loop engineering emerged as models crossed a reliability threshold. Fable 5’s predecessor, Opus 4.8, could not sustain multi-day tasks or self-correct as robustly; Fable 5 handles long horizons and self-validates.
Osmani’s July 15 blog post gave the method a formal label and a history. Prompt engineering was about crafting the right instruction. Context engineering added relevant background. Harness engineering connected models to tools. Loop engineering wraps all three in a recurring, goal-driven cycle. Osmani’s central insight: the engineering work moves from writing prompts to building verification.
What to Do Now
Windows developers who want to experiment with loop engineering can start small. Cherny’s own advice, shared when describing his workflow, is that the single most important step to squeeze maximum performance out of Claude Code is to give it a way to verify its own work.
Here is a practical starting sequence:
- Pick a repeatable, low-risk task. Good candidates: fixing a known CI failure that recurs predictably, monitoring a specific flaky test, or generating a daily summary of pull-request activity.
- Define a goal and verification method. Write a unit test, a linting rule, or a script that must pass. The agent should have no way to claim success unless that independent check succeeds.
- Use existing orchestrators. A GitHub Actions workflow can trigger a Claude Code loop on a cron schedule. Azure DevOps pipelines can do the same. Even a PowerShell script invoked by Task Scheduler can act as a lightweight loop driver when combined with the Claude Code CLI.
- Isolate credentials. Create a fine-grained token that allows only the necessary operations—for example, reading logs and opening draft pull requests, but not merging or deploying.
- Set a budget and a kill switch. Limit API spending and design the loop to stop after a fixed number of attempts or a time window. A manual override should always be available.
- Start with a manual trigger. Before letting the loop run autonomously, invoke it by hand and review the output. Once confident, promote it to a scheduled job.
- Log everything and review periodically. Treat the loop like any other automated workflow: capture its actions, decisions, and artifacts so you can audit when something goes wrong.
- Keep humans in the approval chain for critical steps. Merge and deploy decisions should remain manual until you have extensive, battle-tested confidence in the loop’s verification.
Anthropic’s documentation for Claude Code CLI details how to script these interactions. The loop engineering idea is not locked to Claude Code; any agentic tool that supports test-driven goals can be adapted, but Fable 5’s self-validation and visual comprehension give it an edge for Windows projects that involve UI testing or dashboard analysis.
Outlook
Loop engineering is still an emerging practice, but the tooling is maturing fast. Microsoft’s own Copilot offerings may soon incorporate similar agentic patterns, given the industry direction. For now, Windows developers who automate their most tedious chores with well-fenced loops stand to gain back hours of manual toil. Cherny’s vision of hundreds of agents whirring overnight might sound extreme, but the underlying principle—delegate the routine, verify the output, and keep humans in control of consequences—is a durable one for any team.
As Fable 5 and competing models improve, expect more pre-built loop templates and deeper integration with CI/CD platforms. The real catalyst won’t be a new framework; it will be the organizational trust built by teams who demonstrate that autonomous coding agents can be safe, auditable, and productive when wrapped in the right verification loops.