Agent loops run by themselves; comprehension debt is the new bottleneck

The loop. Loop engineering replaces one-shot prompt writing with a system that keeps prompting an agent until a goal is met. The core cycle is simple: goal, action, observe, adjust. The critical design decision is the stopping condition. "Make checkout faster" is unverifiable, so the agent cannot know when it is done. "Stop when every unit test in the checkout suite passes" gives the loop a measurable end state. That distinction is the practical core of the idea.

The machinery. Around that cycle, the framework adds triggers, context, memory, tools, and verification. Cron schedules or event hooks start the loop. Context must be compacted or progressively disclosed because raw test output piles up. A spine, often markdown files, records what the agent tried and what worked. Tools via MCP or direct APIs let agents execute code, touch files, and query databases. A subagent checks the work. This is system architecture more than prompt writing.

The human role. Despite the "minimal human intervention" claim, the human still writes the goal, documents the intent, and reads the output. Intent debt is the name for what happens when an agent understands what a system does but not why. The transcript's example is sharp: an agent could fix a failing checkout test by relaxing discount code validation, technically correct and commercially wrong. That risk only drops when intent lives in a file like agent.md and stays current.

The debt risk. Unverified code is the first of four pitfalls. A checker subagent is still an agent, so a real person must review what ships. Comprehension debt grows as agents produce more code than humans actually understand. That becomes urgent during production failures and migrations, when someone must reason about the whole system. The fourth pitfall, cognitive surrender, is the most dangerous: a developer accepts agent output without reading it, and then nobody understands the code at all.

The trade-off. The framework's weakness is that it treats autonomy and review as compatible goals. Faster agent loops create more code for the same number of humans to verify, so the review process becomes the bottleneck. The video offers no production data or benchmarks, only a plausible argument. Teams that adopt this model need to decide deliberately how much throughput they are willing to trade for comprehension.

The bottom line. Loop engineering is partly a rebrand of agent orchestration, but it names something worth paying attention to: the limiting factor in AI-generated code is human capacity to set goals, preserve intent, and verify output. That puts the problem on the engineering management side, not the prompt quality side.