Implementing AI Agents with Andrew Ng’s Loop Engineering: A Practical Guide

First Loop: Coding Agent Loop – Self‑Testing Closed‑Loop

Andrew Ng describes the first loop where a Coding Agent receives a specification and a set of evaluation criteria, then autonomously writes code, runs its own test suite, reads logs, fixes failures, and repeats until all criteria pass, eliminating the need for human intervention during minute‑level iterations.

Second Loop: Developer Feedback Loop – Move Beyond AI QA

At the hour‑level, developers shift from hand‑holding AI to guiding higher‑level decisions. They provide contextual insights that AI cannot infer—such as user preferences or product‑specific constraints—by updating specifications and evaluation sets, allowing the AI to focus on code generation while humans steer overall direction.

Third Loop: External Feedback Loop – Real‑World Validation

The slowest loop, spanning days to weeks, releases the partially built product to real users, collects usage data, and feeds it back into the specification, triggering the inner loops to adapt the code accordingly. This external loop grounds the product in market reality and prevents isolated, sandbox‑only development.

Mental Model Shift: From “AI Babysitter” to Pipeline Operator

The core transformation requires abandoning line‑by‑line control and instead designing the input (clear, testable specifications) and the output (automated quality‑check pipelines). The operator’s role becomes defining precise requirements and comprehensive evaluation standards, then letting the AI execute the iterative build‑test‑refine cycle.

Hard‑Core Pitfall Guide: Writing Effective Evaluation Standards

Practical tips include using tools such as Claude Code, Cursor, Codex, or Aider + Git with a --test-cmd "pytest" flag, and creating wrapper scripts that repeatedly run tests until success or a maximum iteration count is reached.

**Supported tools/platforms:**
- **Claude Code / Cursor / Codex** – execute shell commands and testing loops.
- **Aider + Git** – use `--test-cmd "pytest"`.
- **Custom scripts** – simple wrapper to loop tests.

**Sample prompt for the Agent:**
> You are a strict Coding Agent.
> Specification is in spec/login.md.
> After each change, run `pytest tests/login/ -q --cov`.
> Consider the task complete only when all tests pass and coverage ≥85%.
> If failures occur three times consecutively, output a detailed error analysis and stop.

Additionally, maintain a checklist of boundary conditions rather than vague rules; for example, forbid magic numbers in non‑business code and require exception handling after database queries. These concrete engineering red lines define the Agent’s operational boundaries.

By translating informal developer notes into machine‑executable test cases, the evaluation standards become actionable, ensuring the AI operates within well‑defined constraints and reduces reliance on manual oversight.