Loop Engineering: Stop Prompting Your Agents (Real Examples)

Detailed Summary

Introduction (0:00 - 01:40)

Loop engineering is introduced as the successor to prompt, context, and harness engineering. While the term is a buzzword, the core concept involves software that can build itself and achieve product-market fit autonomously. The goal is to shift the human role from 'writing prompts' to 'finding money for tokens' while the system handles growth and development.

The Product Market Fit Loop (01:40 - 03:30)

Using the business 'Inbox Zero' as an example, a high-level loop is described. Instead of a founder waking up to decide on tasks, an agentic loop follows a cycle: build product -> get feedback (via marketing/usage) -> analyze logs/analytics -> improve product -> repeat. This cycle mirrors how human-led businesses operate but aims for full autonomy.

Designing Loops vs. Prompting (03:30 - 05:00)

Industry experts Boris Cherny and Peter Steinberger suggest that developers should stop prompting agents directly and start designing loops that prompt agents. This involves creating an orchestrator or 'scaffolding' that runs on a cron job (e.g., every 24 hours) to perform minor improvements and self-checks without human intervention.

The Six Pieces of a Loop System (05:00 - 09:00)

Building a robust loop requires six key components:

• Automation: Scheduling tasks to run independently.
• Work Trees: Allowing multiple agents to work on separate branches in parallel without conflicts.
• Skills: Providing agents with specific project knowledge via markdown files (e.g., skills.md).
• Plugins/Connectors: Using Model Context Protocol (MCP) or CLIs to give agents access to databases, logs, and communication tools like Slack.
• Sub-agents: Implementing a 'peer review' system where one agent writes code and another independently verifies it.
• Memory: Storing task lists and progress in markdown files or linear boards on disk, as LLMs lose context between runs.

Running Loops in Claude Code and Codex (09:00 - 10:10)

Practical implementation is now built into modern tools. Claude Code features a sloop command for recurring checks and a goal command for task-oriented loops. These tools allow agents to monitor repositories for issues and attempt to resolve them on a schedule.

Peter Steinberger's Scale and OpenClaw (10:10 - 12:30)

Peter Steinberger manages OpenClaw using massive token leverage, spending approximately $1.3 million monthly on tokens. This allows a tiny team to manage over 92,000 pull requests. While most businesses cannot afford this scale, the principle of using 'auto-review' bots to handle low-risk PRs demonstrates the power of high-leverage loop engineering.

Real-World Loop Ideas (12:30 - 16:40)

Several actionable loop ideas are presented:

• Trend-Following Loop: An agent that monitors Reddit/GitHub for trends and suggests repo improvements daily.
• Error-Fixing Loop: Automatically pulling Sentry or Axiom logs and generating PRs to fix production bugs.
• Eval-Driven Loop: Using a 'goal' to refine prompts and tools until an agent reaches a specific performance score (e.g., 90% on a benchmark).
• Review Loop: A cycle where an agent implements a feature, a bot reviews it, and the agent fixes the issues until the code is ready for human merging.

Conclusion (16:40 - 17:20)

The ultimate business loop is an AI that asks 'how can I improve the business today?' across marketing, sales, and product. Viewers are encouraged to identify repetitive tasks in their own workflows that can be converted into autonomous loops.

Notable Quotes

"In 2026, you don't prompt anymore. Your software should be able to build itself and achieve product market fit on its own." — Dimmitro (quoted by Elie)

"I don't prompt Claude anymore. I have loops running that prompt Claude and figuring out what to do. My job is to write loops." — Boris Cherny, Head of Claude Code at Anthropic

"The agent forgets, the repo doesn't." — Elie