What Is Loop Engineering and Why It’s the Next Step for AI Coding Agents

Definition

Loop Engineering, introduced in June 2026, extends the Prompt → Context → Harness engineering roadmap by adding an automation control structure that decides when and how to prompt AI coding agents, eliminating manual prompting.

Three‑layer progression

Prompt Engineering : Optimises how to write a single instruction; the work unit is the manual dialogue you type.

Context Engineering : Optimises what documents, history, or tool definitions to place in the model’s context; the work unit is the pre‑conditions for one model response.

Harness Engineering : Optimises the runtime environment for a single agent; the work unit is a complete agent conversation.

Loop Engineering : Optimises who decides the prompt, when to prompt, and whether the result is accepted; the work unit is a self‑running loop that spans multiple conversations.

Six components required for a loop

Automations – Scheduled tasks that trigger the loop and discover work items.

Worktrees – Separate directories so parallel agents do not interfere with each other.

Skills – Project knowledge captured in SKILL.md (e.g., conventions, build steps, known pitfalls).

Plugins/Connectors – MCP‑based adapters that let agents interact with real tools such as issue trackers, databases, CI systems, or Slack.

Sub‑agents – Distinct agents for code generation (maker) and code review (checker) to avoid self‑deception.

Persistent memory – Disk‑based state files (e.g., TODO.md, AGENTS.md) that survive across agent runs.

Tool‑specific terminology

Automations : Codex – “Automations” tab (project + prompt + frequency); Claude Code – /loop command, hooks, GitHub Actions.

Long‑term goal : Codex – /goal (available after version 0.128.0); Claude Code – built‑in validator with a separate model that decides completion.

Worktrees : Codex – per‑thread worktree; Claude Code – git worktree, --worktree, isolation worktree.

Skills : Both platforms use SKILL.md files.

Connectors : Codex – MCP connectors + plugins; Claude Code – MCP servers + plugins.

Sub‑agents : Codex – .codex/agents/ TOML files; Claude Code – .claude/agents/ + agent teams.

Persistent state : Codex – Markdown file or Linear; Claude Code – Markdown ( AGENTS.md) or Linear via MCP.

/goal primitive

The /goal primitive lets a loop decide when a task is finished by delegating the “completion” judgment to a smaller maker‑checker model rather than the code‑writing model.

# Claude Code: daily 9 AM triage
/loop "Read yesterday’s CI failures and open issues, write to TODO.md, fix quick‑win items" --schedule "0 9 * * 1-5"

# Claude Code: run until condition met
/goal "All tests under test/auth pass and lint is clean"

# OpenAI Codex: persistent goal
codex /goal "Migrate billing module to new pricing API, keep all existing tests green"

End‑to‑end loop example

Automations (trigger each weekday morning)
  │
  ▼
Call Triage Skill (read CI failures, open issues, recent commits)
  │
  ▼
Write progress to TODO.md (persistent memory)
  │
  ▼
For each worthwhile finding:
  ├─ Open an isolated worktree
  ├─ Maker sub‑agent writes code
  └─ Checker sub‑agent reviews code
  │
  ▼
Connectors open PR + update Issue
  │
  ▼
Schedule next run for tomorrow morning

The state file ( TODO.md or AGENTS.md) records what has been tried, what succeeded, and what remains, allowing the loop to continue from the previous day’s progress.

Problems loops do not solve

Verification remains the developer’s responsibility. Even with maker‑checker separation, the “done” declaration is not a proof; human review of the final code is still required.

Code‑understanding debt can accelerate. Loops generate code faster than a developer can read it, widening the gap unless the developer continuously audits the output.

Cognitive surrender becomes an easy failure mode. When a loop runs smoothly, developers may accept its output without scrutiny; the loop itself is neutral, but design choices determine whether it amplifies insight or error.

Getting started

Apply loops to repetitive, verifiable tasks (e.g., daily CI triage).

Retain manual control for judgment‑heavy, one‑off tasks.

Start with a simple scheduled task that reads CI failures each morning and writes them to a Markdown file for review. After observing loop behaviour, gradually add sub‑agents and automated PR creation.

Monitor token consumption: each loop trigger can consume many tokens, and goals with built‑in validators consume more. Begin with low frequency, observe billing for a few days, then scale up.

References

Addy Osmani – Loop Engineering: https://addyosmani.com/blog/loop-engineering

Lushbinary guide: https://lushbinary.com/blog/loop-engineering-ai-coding-agents-guide

Addy Osmani – Agent Harness Engineering: https://addyosmani.com/blog/agent-harness-engineering

Outcome School – AI Agent Loop: https://outcomeschool.com/blog/ai-agent-loop

Twitter trend “Developers Shift to Loop Engineering”: https://x.com/i/trending/2063869701080437219