Are Prompts Becoming Obsolete? A Deep Dive into Loop Engineering

Loop Engineering is the next step after Harness Engineering, aiming to eliminate manual prompting of coding agents by designing a system that automatically triggers and manages agents.

1. Why Loop Engineering Appears

It emerged as a response to the inefficiency of manually writing prompts for agents; experts like Peter Steinberger and Boris Cherny now run loops instead of hand‑prompting.

2. The Five Core Modules + Memory

Automations : scheduled tasks that automatically discover and triage work.

Worktrees : isolated checkouts (via git worktree) that prevent file‑level conflicts when multiple agents run in parallel.

Skills : reusable project knowledge stored in SKILL.md files, reducing repeated intent specification.

Plugins and Connectors : integrate agents with existing tools (issue trackers, databases, Slack, etc.) using MCP‑based connectors.

Sub‑agents : separate agents for generation and verification, preventing self‑bias.

Memory is a persistent store (e.g., a Markdown file or Linear board) that records completed tasks and next steps, ensuring long‑running agents retain state.

3. How Claude Code and Codex Implement Loops

Both products embed the five modules. Automations can be created via a UI, selecting prompts, frequency, and execution context (local checkout or worktree). Skills are invoked automatically when task descriptions match. Plugins/Connectors allow loops to open PRs, update tickets, and post to Slack. Sub‑agents are defined in .codex/agents/ or .claude/agents/ with TOML or configuration files, enabling separate generation and review agents.

4. Human Responsibilities

Even fully automated loops require human verification; the “completion” signal is only a declaration, not proof. Engineers must review generated code to avoid quality decay and maintain understanding of the codebase.

5. Designing Effective Loops

Engineers should balance automation with oversight, using loops to accelerate repetitive tasks while retaining judgment for critical decisions. Proper use of worktrees, skills, and sub‑agents mitigates token costs and conflict risks.

Overall, Loop Engineering represents a shift in how developers collaborate with AI agents, moving from prompt‑centric workflows to system‑centric loops that automate discovery, execution, and verification.