Summary of Key Points
This article focuses on the recently popular Loop Engineering in Silicon Valley: a system that enables AI programming agents to complete tasks automatically. Instead of manually writing instructions, users create an “automatic assembly line” that allows the AI to find tasks, assign them, verify results, and track progress on its own, in a continuous cycle. The first half of the article explains the definition, core functions, and components of Loop Engineering. The second half reveals industry insights from a closed-door discussion: the potential of Loop Engineering is often overestimated, while the complex interconnections between tasks (topology) are underestimated. The discussion also covers aspects such as multi-agent collaboration, the importance of feedback, trends in environment construction, and the practical challenges in implementing these systems.
1. What is Loop Engineering? – From “Manually Guiding AI” to “Building an Automatic Assembly Line”
In the past, when using AI for programming, you had to write detailed instructions (e.g., “Fix this bug for me”) step by step. With Loop Engineering, the process is reversed: you build a system that allows the AI to complete the entire task itself, involving five steps:
1. Finding Tasks: The AI automatically scans code repositories to identify unresolved tests, issues, or deteriorating metrics.
2. Assigning Tasks: Each task is assigned its own independent “working copy” to prevent multiple AI instances from modifying the same file simultaneously.
3. Verification: Another AI acts as a “referee” to check the first AI’s work (since self-assessment tends to be biased).
4. Tracking Progress: The system records what was done and the results in files or databases to prevent the AI from forgetting.
5. Scheduling: A scheduled time (e.g., every morning) is set for the system to run automatically, without the need for manual intervention.
This system is built using six existing components: tools for automatically finding tasks, independent working copies, a project guideline manual, interfaces to connect with daily tools (such as Slack), sub-AIs responsible for different tasks (coding and reviewing), and external storage systems (e.g., markdown files). For example, Addy’s morning routine involves the system automatically scanning the code repository, assigning bugs to AI agents for repair, and having another AI review the results; only difficult cases are then handed over to humans.
2. Why has Loop Engineering Suddenly Become Popular? – Improved Model Capabilities and Availability of Necessary Components
Previously, such systems were ineffective because AI was limited in its capabilities, requiring detailed instructions and manual script maintenance. The current popularity is due to two main reasons:
1. AI models have surpassed certain thresholds: They can now complete complex tasks without human supervision, and scheduled operations have become standard features.
2. Components are readily available: Tools like Claude Code and Codex already provide the necessary functionalities for automatic task discovery and independent work management.
For example, Stripe’s Minions system automatically merges over 1,300 pull requests (PRs) per week, all done by AI. This would be impossible with manual instruction. Another example is Anthropic’s demonstration: simple instructions are insufficient to run a game; however, using Loop Engineering, the game’s appearance and functionality can be adjusted through repeated iterations, although it requires more time and resources.
3. The Challenges of Loop Engineering
Loop Engineering does not solve all problems; it merely exacerbates some existing issues in AI programming:
1. Intent Debt: Without clear project guidelines, AI may make random changes (e.g., modifying core code), so these rules must be documented for the AI to follow.
2. Verification Debt: Even with automated verification, human review is still necessary before the system goes live.
3. Understanding Debt: The faster AI codes, the more code users don’t write or understand, making future maintenance more difficult.
4. Cognitive Reluctance: Users may become dependent on AI’s recommendations and lose their ability to make independent decisions.
5. Orchestration Costs: Running multiple AI tasks in parallel incurs additional costs (e.g., token fees), which need to be carefully managed.
Peter Steinberger acknowledges that using multiple Loop Engineering systems is a temporary solution until models become faster.
4. Critical Insights from the Closed-Door Discussion: Overestimating Looping, Underestimating Topology
In private industry discussions, experts are more skeptical. The main issue is that repetitive processes (looping) are overemphasized, while the complex structure of tasks (topology) is underestimated:
- Loop-based tasks: These are easier to manage with Loop Engineering (e.g., fixing tests), but real-world tasks often involve multiple interrelated subtasks.
- Real-world tasks combine both: You need to repeatedly optimize specific modules (looping) while also advancing the overall structure of the system. For example, developing an app requires optimizing login functions and coordinating payment and social features.
- The difficulty of topology: It’s hard to define a “reasonable product structure,” which makes it easier for AI to exploit loopholes, leading to preference for simpler looping tasks.
Another point is that long-running processes (e.g., hours-long AI operations) are not necessarily more complex. Manufacturers may highlight these, but real users typically use AI for shorter periods (e.g., 10 minutes). Human work is often parallel and interrupted, which is difficult to replicate with single-loop systems.
5. Practical Challenges in Implementation: Far from Replacing Humans
Although Loop Engineering sounds promising, there are significant gaps before it can fully replace human tasks:
1. Complex Tasks: AI struggles with complex tasks like consulting-level PPTs, which require high-quality, structured information.
2. Idle Time: Users may wait for long periods while AI performs simple tasks; predicting task duration and providing updates (e.g., “Wait 3 minutes for more guidance”) is challenging.
3. Self-Improvement: There’s no theoretical proof that AI can improve itself; current “improvements” are based on trial and error. Better models may only make a few mistakes, while others may need multiple attempts.
In conclusion, there’s no need to panic (FOMO). By trying Loop Engineering yourself, you’ll realize that current models are still far from being perfect, and real user needs are more complex than imagined.
6. Managing AI Like Managing Employees
The article concludes with a fitting analogy: Managing AI is similar to managing employees. If you assign tasks without providing feedback, the results will not meet your expectations. We’re still far from having AI take over entire workflows, but we have made significant progress in terms of environment, data, feedback mechanisms, and evidence of self-improvement.
(End of Article)