Loop Engineering

Loop engineering advocates no longer directly prompting coding agents, but instead designing loops that prompt agents and defining quantifiable success criteria for agents to self-evaluate and continue running.

Context Engineering focuses on how to precisely manage the context window fed into the model, Harness Engineering leans toward the infrastructure layer, and both appeared around the same time as Loop Engineering.

The core engineering concerns of Loop Engineering include loop granularity design, error recovery mechanisms, observability, and cost control.

Loop Engineering does not replace Prompt Engineering, but rather encompasses and makes good use of it.

There is a risk that a single Agent handling tasks may go off track (hallucinate), potentially misjudging an order and issuing an erroneous full refund.

The "Loop" in Loop Engineering and the "Loop" in Agent Loop are not the same concept: the former is a loop of multiple Skill calls and multi-Agent collaboration, while the latter is a loop of tool calls and reasoning.

Loop engineering is particularly effective in complex task scenarios that are difficult to complete in a single step, such as code generation, automated project deployment, and multi-step process orchestration.

The core of Loop Engineering is a role shift: people move from being operators inside the loop to being system designers outside it, designing systems that can discover tasks, dispatch them, verify them, store state and launch the next round on their own.

A Loop round consists of five actions—discovery, handoff, verification, persistence and scheduling—supported by six components: automation, workflows, skills, connectors, sub-agents and external memory.

The three disciplines of Loop Engineering: sample and read Loop output every day and explain every change; first set per-run/per-day budgets and a maximum retry count as circuit breakers before talking about automation; always keep a human intervention entry point.

Loops carry four types of invisible bills: validation debt, understanding rot, cognitive surrender, and Token explosion, and these four reinforce each other.

Five kinds of runaway Loops correspond to five skipped actions: skipping validation leads to the nodding Loop, skipping persistence leads to the forgetful Loop, skipping scheduling leads to the manual Loop, skipping discovery leads to the blind Loop, and skipping handover leads to the tangled Loop.

In a Loop, the judge should be structurally swapped: one Agent writes and another Agent with different instructions reviews, even switching the underlying model, similar to the separation of generator and discriminator in a GAN.

The typical Loop Engineering workflow is: select a task, write code, run tests, automatically fix errors, and consider it done only when all tests pass, focusing on whether the entire loop can converge on its own.

ReAct describes the "thought-action-observation" reasoning structure within a single round, while Loop Engineering focuses on organizing these rounds into an engineering closed loop with acceptance criteria, automatic retries, and controllable cost.

The video only gives definition-level descriptions of the four concepts and does not elaborate on implementation details, failure modes, or differences between frameworks.

Loop Engineering requires verifiable termination conditions (such as tests passing) and cannot rely on the model judging for itself that it is "done"; loops without acceptance criteria easily exhaust the budget or drift from the goal.

Loop Engineering refers to outsourcing the routing work of feeding prompts back and forth between multiple agents to Python or TypeScript scripts for automatic execution.

The loop allows agents to recover from partial failures, but does not understand why the failure occurred, whether it has occurred before, or whether the system needs to change to prevent recurrence.

The author distinguishes three layers: prompt engineering concerns optimizing a single response, Minds concerns shaping the agent itself, and Loop Engineering concerns shaping how the agent operates over time.