Loop Engineering Explained: Evolution, Six Core Components, and Control Theory

01 Prompt → Context → Harness → Loop

In 2024 the dominant way to work with AI was Prompt Engineering: a single prompt had to contain background, rules, examples, and constraints, which limited scalability because the model could only guess what was omitted. By early 2025 the community realized the bottleneck was insufficient context, giving rise to Context Engineering, which enriches the model’s view with system prompts, few‑shot examples, RAG retrieval, and structured inputs. However, even with perfect context, complex tasks still required multi‑step execution, tools, and intermediate observation, leading to the emergence of Harness Engineering. Harnesses equip agents with tools (shell, file system, API connectors, sandbox) and retry mechanisms so they can complete multi‑step sessions.

Despite this progress, the human remained the bottleneck: agents could read, write, test, and open PRs, but developers still had to manually trigger the loop each morning, copy CI logs, feed them to the agent, review results, and approve merges. Loop Engineering replaces the human‑in‑the‑loop with an automated system that can trigger, validate, remember state, and decide whether to continue or stop.

02 What Loop Needs

The industry‑standard Loop consists of six components, illustrated with a CI‑failure scenario:

1. Automations (Trigger) : Detect CI failure (via cron, webhook, or heartbeat) and define the work target.

2. Worktrees (Isolation) : Run fixes in a separate Git worktree to avoid interference between concurrent tasks.

3. Skills (Knowledge) : Load project‑specific knowledge files (e.g., AGENTS.md, CLAUDE.md, SKILL.md) so the agent does not have to relearn conventions each session.

4. Connectors (Actuation) : Use tools and APIs (GitHub, Linear, Slack, databases) to create PRs, close tickets, or send notifications.

5. Sub‑agents (Sensor) : A second agent validates the result; the primary agent never validates its own output.

6. Memory/State : Persist the outcome of each run (success, failure, attempted steps) so the next iteration knows where it left off.

When these components are wired together, the Loop automatically detects a red CI, isolates a worktree, applies known skills, fixes the code, creates a PR, validates with a sub‑agent, records the state, and repeats until the CI passes.

03 Loop Essence: Control Theory

Most discussions treat Loop Engineering as a choreography problem, but its deeper nature is control theory (cybernetics). A closed‑loop system has three roles: Controller (decides actions), Actuator (executes actions), and Sensor (measures deviation). In the CI example, the goal is “all tests pass”; the agent’s code changes are the execution; the test suite provides the error signal; the controller decides the next prompt based on that signal.

Control theory explains why a Loop differs from a static cron job: the decision logic resides in the LLM, which can adapt its actions based on observed deviations rather than following a hard‑coded if‑else.

The six Loop components map onto control theory:

Skills and Memory support the Controller.

Connectors and Worktrees support the Actuator.

Sub‑agents serve as independent Sensors.

Automations provide the start‑signal for the closed loop.

Three possible Loop outcomes are described:

Convergence to a correct state : the agent reaches the goal and verification is trustworthy.

Convergence to an incorrect state : the loop stops because the sensor gave a false positive.

Divergence : repeated attempts push the system farther from the goal until a hard limit stops it.

Because the sensor determines convergence speed, a rich sensor that reports detailed failure information (which test failed, which diff caused it) dramatically narrows the search space compared to a binary pass/fail sensor.

04 When to Build a Loop

Adopting a Loop requires that you can automate a reliable check that rejects bad output without human oversight. The task must be repeatable (e.g., a CI failure that occurs at least weekly) and the token budget must accommodate repeated context loading and retries. Objective pass/fail criteria are essential; subjective goals like “make the code more elegant” lack measurable sensors and lead to false convergence.

If you meet these conditions, start by writing Skills (explicit intent and conventions), then design a precise Sensor (detailed verification), and finally add the Automation trigger (cron or webhook). The Loop amplifies the quality of the sensor you provide.

附录、参考资料
1. Loop Engineering: The New Way to Use Claude Code & Codex (https://medium.com/towards-artificial-intelligence/loop-engineering-the-new-way-to-use-claude-code-codex-c55dd65ecc61) — Addy Osmani
2. Loop Engineering Is NOT What Everybody Thinks It Is (https://medium.com/@agentnativedev/loop-engineering-is-not-what-everybody-thinks-it-is-6719a0f4f83f) — Agent Native Dev
3. How Claude Code, Codex, and Cursor Do Loop Engineering (https://medium.com/ai-all-in/how-claude-code-codex-and-cursor-do-loop-engineering-28b444968673)
4. Loop Engineering (https://medium.com/@cobusgreyling/loop-engineering-62926dd6991c) — Cobus Greyling
5. Why Is Loop Engineering Trending (https://medium.com/generative-ai/why-is-loop-engineering-trending-2acb7029af0c)
6. Loop Engineering Is Replacing Prompt Engineering (https://medium.com/coding-nexus/loop-engineering-is-replacing-prompt-engineering-why-the-best-developers-dont-prompt-ai-anymore-639c944be3ee)
7. What is Loop Engineering? How it is different than Harness Engineering (https://medium.com/gitconnected/what-is-loop-engineering-how-it-is-different-than-harness-engineering-0e764f373fb1) — Akshay Kokane
8. Loop Engineering Is Here. Most of You Should Not Build One Yet (https://medium.com/@alirezarezvani/loop-engineering-is-here-most-of-you-should-not-build-one-yet-part-1-d56f63acda00)
9. How To Build a Claude Loop Engineering Better Than 99% of People (https://medium.com/data-science-collective/how-to-build-a-claude-loop-engineering-better-than-99-of-people-3ab8701d176c)