From Prompt Engineering to Loop Engineering: AI Agents Now Run Continuous Workflows

From Prompt to Work Process

The author began using Codex to create presentations. Initially, each slide required a new prompt, with the human directing every step. After several iterations, Codex started running for over an hour without new prompts, reusing earlier requirements and context to complete the file. This behavior mirrors the emerging concept of Loop Engineering : organizing goals, context, rules, tools, and checks into a self-sustaining process rather than relying on single prompts.

Loop Engineering vs. Prompt Engineering

Traditional Prompt Engineering focuses on crafting one instruction for one action. Loop Engineering asks how to structure a continuous workflow. The author illustrates with an R&D scenario: an automated loop that checks CI failures, spawns a worktree, lets an agent analyze and fix, runs tests and lint, invokes a reviewer agent, opens a PR on success, or logs failure for the next iteration or human handoff. This is a full automated work process, not a single prompt.

Conditions for a Usable Loop

Specific goals: "Optimize this project" is too vague; "make test/auth pass, tsc --noEmit zero errors, npm run lint zero violations" is executable.

Stable context: Each agent start is like a new hire; without project docs, code structure, team rules, test methods, and delivery standards, the agent repeats questions or follows wrong assumptions.

Tools integrated into real workflows: Chat-only AI is limited. Connecting GitHub, CI, docs, task systems, and file systems puts the agent on the actual project site.

Verification mechanisms: A loop without result checking becomes an automatic error generator. It must know success, failure, when to continue, and when to hand off to humans.

Codex's Evolution: Loops Emerging from the Tool

The author did not pre-design a complex pipeline for the PPT task. By continuously providing requirements, feedback, and confirmations (page style, title handling, element placement, occlusion rules, output format), Codex began remembering and applying these constraints across subsequent steps. It maintained style consistency, adjusted layouts based on prior feedback, and finally synthesized the deck, added watermarks, and produced an overview image. This shows loops can emerge from the agent's own memory and context, not only from human-designed loops. The human role shifts from step-by-step operator to direction-setter, rule-definer, and calibrator.

Why This Matters: Rising Abstraction Level

AI programming abstraction is climbing: from hand-written code, to IDE autocomplete, to prompt-driven agents, to goal/rule/check-defined continuous agent work. This is an engineering problem—making a process repeatable, collaborative, verifiable, and improvable, with AI agents as executors. The pattern extends beyond code: a PPT can be decomposed into theme definition, material gathering, structure generation, layout design, page generation, style unification, and export. If AI remembers earlier decisions and auto-handles later steps, the loop prototype exists. Many future tasks may shift from "question-answer" to "continuous processing."

Risks: Loops Must Not Run Unchecked

A bad prompt fails once; a bad loop fails repeatedly. Vague goals cause endless trial; weak checks treat errors as success; stale project knowledge leads to compounding mistakes; excessive tool permissions amplify small issues. Loop Engineering therefore tests three abilities: defining clear goals, designing reliable checks, and ensuring the system stops, logs, and hands off on failure. This does not eliminate engineers; it moves them from per-step operators to system designers and result calibrators.

Three Engineering Layers

Prompt Engineering: How to clarify the current action to AI.

Context Engineering: What background to give AI so it understands the project and task.

Loop Engineering: How to keep AI working continuously in a verifiable process.

The author's Codex experience suggests AI agents are no longer passive responders; they start looping tasks autonomously using memory, context, and tool capabilities. The highest value will belong to those who can wire goals, tools, context, checks, and feedback together—and know how to calibrate that loop.