Loop Engineering in 9 Minutes

Detailed Summary

Why Loops Beat Prompts (0:00 - 0:43)

• The creator of Claude Code recently noted a shift from prompting Claude to building loops that perform the majority of his job.
• This concept represents a spectrum of complexity, moving from hands-on interaction to autonomous systems.
• The release of Fable 5 and insights from Lance Martin highlight how loops are becoming the primary way to interact with advanced AI models.

Loops Spectrum and Examples (0:43 - 1:18)

• Andrej Karpathy’s "Auto Research" project is a prime example, where a model trains itself by running experiments at set intervals.
• The system evaluates what works and what doesn't, progressing only the successful experiments.
• This "parameter golf" approach can be applied to both programming and broader non-technical tasks.

Goals: Long Running Agents (1:18 - 1:56)

• A "goal" is defined as a long-running task that doesn't require constant re-prompting.
• Tools like Codex and Claude Code allow users to simply type goal followed by a natural language description.
• The harness has built-in mechanisms to carry out the task autonomously until the objective is reached.

Claude Code Slash Loop (1:56 - 3:34)

• The /loop feature in Claude Code allows for interval-based execution using natural language (e.g., "loop every 5 minutes").
• This is particularly useful for exploratory or green-field projects where the AI acts like an "ambitious junior engineer."
• Users can set deterministic cron jobs within an LLM thread, such as running a loop for 5 minutes every hour for a 3-hour duration.

When to Use Goal vs Loop (3:34 - 4:18)

• Goals are preferred for bounded, complex tasks that can run for multiple days, such as building parsers for complicated documents.
• The effectiveness of a long-running goal increases significantly when the AI has a way to verify its work, such as running unit tests.
• Bounded tasks with verifiable outcomes are the ideal use case for multi-day autonomous agents.

IDE Automations Tab (4:18 - 5:11)

• Modern IDEs like Cursor, Codex, and Claude Code now feature an "Automations" tab, usually in the top-left corner.
• This interface allows users to manage active and paused automations, moving beyond one-off prompts to consistent background workflows.
• These tools are designed to handle the "minutia" of daily work that would otherwise require manual intervention.

Daily Inbox to Linear (5:11 - 6:00)

• A practical application involves an automation that scans a YouTube email inbox and triages requests onto a Linear board.
• The agent checks the inbox and the board daily to identify and prioritize high-action items.
• Automations can also be used to maintain project documentation, such as creating agent.md files or updating skill logs weekly.

Security Scans and Memory (6:00 - 6:44)

• Loops can be scheduled to perform security vulnerability scans on a project at a specific cadence.
• An emerging concept is "synthetic memory," where an agent "dreams" or reflects on the day's events to create a coherent summary.
• This process creates a representation of data that the model can programmatically load into its context window later.

Continual Learning Future (6:44 - 7:49)

• Continual learning allows systems to move beyond their static pre-training data by exploring and building on new tasks in the real world.
• This approach enables AI to progressively get smarter and more useful across various domains over time.
• Most repetitive computer-based tasks are candidates for being offloaded to these automation loops.

Human in the Loop Tips (7:49 - 8:12)

• Automations do not need to be end-to-end; humans should remain in the loop for final approvals.
• For example, an AI can draft emails, but the human should review and send them to ensure context is correct.
• Over time, the AI can even improve its own prompts based on human feedback and corrections.

Wrap Up (8:12 - 08:49)

• The industry is moving toward scaffolding automations and thinking at a higher level of system management.
• Viewers are encouraged to consider if they are spending less time prompting and more time engineering loops.

Notable Quotes

"I don't prompt Claude anymore. What I mostly now use are loops. I create loops, they do the rest of my job." — Creator of Claude Code (referenced)

"Think about it handing it off to a really ambitious junior engineer... code just generally is becoming cheap, you can do much more ambitious things." — Speaker

"If something that you are repeatedly doing and it is on a computer, there is a pretty good chance that you're going to be able to offload a lot of these tasks to these types of automations." — Speaker