Only the best are using them...

Too Long; Didn't Watch — Summary

Loop engineering is the next evolution of software development where engineers design autonomous systems that self-correct and iterate until a specific goal is met, moving beyond manual prompt-by-prompt interaction.

Main Takeaways

• Shift from Prompting to Designing: The industry is moving from 'vibe coding' (manual prompting) to designing loops that manage agents autonomously.
• The Two Pillars of Loops: A functional loop requires a specific trigger (action, schedule, or human) and a verifiable goal (deterministic or LLM-judged).
• High Barrier to Entry: Currently, only the top 1% of engineers at companies like OpenAI and Anthropic use these techniques due to extreme token costs.
• Automation vs. Looping: Unlike simple automation, a loop includes internal decision-making to evaluate if a goal has been reached before terminating.
• Future of Engineering: Software development is transitioning into the creation of 'software factories' that may eventually lead to recursive self-improvement.

Detailed Summary

Introduction (0:00 - 01:08)

Recent viral discussions from industry leaders Peter Steinberger and Boris Cherny have highlighted a shift in AI coding strategy.

• The 'new meta' involves moving away from manual prompting of models like Claude.
• Top engineers are now focused on writing loops that allow AI to figure out tasks independently.
• This strategy represents the absolute frontier of modern software engineering.

Defining the Loop (01:09 - 02:38)

Traditional agentic engineering requires constant human intervention, whereas loops automate the iteration process.

• A loop is defined by specifying an end state or goal rather than a step-by-step instruction.
• Goals must be verifiable, either through deterministic tests (code passing) or non-deterministic LLM evaluation.
• This process mirrors reinforcement learning, where the agent receives a 'reward' or signal upon successful completion.

Practical Implementation and Triggers (02:39 - 05:43)

Loops can be integrated into modern tools like Cursor and managed through specific triggers.

• Action Triggers: Events like opening a Pull Request (PR) can trigger an agent to review, fix, and commit code automatically.
• Schedule Triggers: Also known as cron jobs, these run loops at set intervals (e.g., every 30 minutes).
• Human Triggers: A developer provides a full specification and initiates the loop to run until completion.
• The complexity of a loop increases when the goal is 'amorphous,' such as building a feature from a broad spec.

Tools and Sponsorship (05:44 - 07:05)

As loops produce more code, tools are needed to handle rapid deployment and storage.

• here.now is highlighted as a tool that gives agents the ability to publish directly to the web.
• Agents can install skills to manage private storage and custom URLs.
• This infrastructure supports the high-velocity output generated by autonomous loops.

Claude Code and Loop Examples (07:06 - 08:11)

Specific AI tools are already incorporating looping as a native feature.

• Claude Code features a /loop command to run prompts on a recurring basis.
• Example: A loop can be set to compare current progress against a spec.md file every five minutes until parity is reached.
• This removes the need for the human to check in after every small code change.

The Challenges: Complexity and Cost (08:12 - 10:22)

Despite the potential, loop engineering faces significant practical hurdles.

• Difficulty: Designing a 'code factory' that doesn't fail or hallucinate is technically demanding.
• Expense: Looping consumes a massive amount of tokens; some top engineers report monthly bills exceeding $1.3 million.
• Bifurcation: There is a growing gap between elite engineers with 'infinite' token budgets and the rest of the industry.

The Future: Software Factories (10:23 - 12:58)

The role of the engineer is fundamentally changing from a writer to an architect.

• Engineers will eventually stop writing prompts and start designing autonomous factories.
• The distinction between automation and loops lies in the 'decision' step—the ability of the system to judge its own progress.
• The ultimate horizon is recursive self-improvement, where AI begins to set its own goals and design its own improvements.

Notable Quotes

"I don't prompt Claude anymore. I have loops that are running. They're the ones that are prompting Claude... My job is to write loops." — Boris Cherny

"You shouldn't be prompting coding agents anymore. You should be designing loops that prompt your agents." — Peter Steinberger

"We as engineers will gradually and then suddenly no longer even be writing prompts to our agents... We will be designing these factories." — Matthew Berman