Prompt Engineering Is Dead—Enter Loop Engineering: Is AI Coding Making Work Easier or Harder?

01 From Prompt Engineering to Loop Engineering

Most developers initially think AI that writes code will make work easier, but Google Cloud AI director Addy Osmani argues the opposite: mastering this approach is harder than writing prompts. He illustrates the shift by comparing a fully present engineer with one who only presses a start button, coining the term “Loop Engineering”.

Claude Code lead Boris Cherny says he no longer gives direct prompts to AI; instead he writes loops that drive the AI, summarizing his work as “my job is writing loops.”

02 What Loop Engineering Actually Is

Loop Engineering means you no longer manually steer an agent each step; instead you turn the “discover‑execute‑check‑record‑continue” chain into an autonomous mini‑system. The focus is on a closed‑loop rather than simple automation.

To be a true loop it must:

Start automatically (scheduled, event‑driven, or upon task completion).

Isolate parallel agents (e.g., using separate worktrees) to avoid conflicts.

Encode project knowledge as persistent “skill” rather than ad‑hoc prompts.

Access local resources (issue trackers, CI, PRs, databases) to act, not just advise.

Separate execution from verification, possibly using different models for review.

Maintain memory of progress (e.g., markdown logs, boards) so the AI does not repeat work.

03 A Complete Loop Looks Like

Addy breaks the loop into five building blocks plus a memory mechanism, which translate into six questions the loop must answer:

Who wakes the loop? Scheduling or event triggers decide whether the system runs autonomously or requires manual initiation.

How to handle parallel agents? Isolation via worktrees prevents file‑level conflicts.

How does the AI know the project’s conventions? Persistent “skill” documentation supplies startup rules, naming conventions, and known pitfalls.

Can it access local data? Integration with issue systems, CI, test environments, and PR tools lets the AI act, not merely suggest.

Who validates the results? A separate verification agent (or model) reviews outputs, mirroring the practice of not self‑reviewing code.

How does it remember progress? External storage (markdown, boards, databases) records successes, failures, and pending items.

04 Costs and Boundaries of Loop Engineering

Loops consume tokens repeatedly for context, trial‑and‑error, and verification, potentially inflating costs if the task isn’t worth repeated runs. Loops also have clear boundaries: they can drive processes but cannot assume responsibility; unchecked loops may produce silent errors and erode developer understanding.

05 Which Work Is Suitable for Loops

While code is the most obvious domain—thanks to clear feedback from tests and logs—any repetitive, stable workflow with verifiable outcomes can benefit, such as content curation, research aggregation, or product analysis. The key criteria are repeatability, stability, and partial automatic verification.

06 The Collaboration Model Has Evolved

Loop Engineering is a transient term; the underlying shift is permanent: as AI handles longer chains, human‑AI collaboration moves from step‑by‑step prompting to designing self‑sustaining loops. Future competition will focus on who designs better loops—scheduling, validation, recording, and termination—rather than who writes better prompts.

Ultimately, loops do not eliminate work; they relocate human effort from repetitive actions to oversight, judgment, and strategic control.