循环工程

AI development has evolved through four stages: single large model calls, prompt engineering, agents and tool calls, and loop engineering.

The technical implementation of persistent agents typically relies on core components such as persistent memory, tool-calling capabilities, and task-scheduling mechanisms.

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.

Each of the three loop layers depends on a reliable context supply: the execution layer needs the current task state, the verification layer needs the original requirements to compare against the output, and the improvement layer needs complete traces.

The "loop engineering" framework proposed in the LangChain Deep Agents course breaks the feedback mechanism down into three levels: the agent loop (execution layer), the verification loop (quality-control layer), and the cross-run improvement loop (evolution layer).

The three layers of loops do not all have to be implemented; developers can selectively stack loop levels according to task complexity and fault-tolerance requirements: the execution layer is the foundation, the verification layer handles quality control, and the cross-run improvement layer targets continuous optimization.

A good goal must be verifiable and have a stopping condition; for example, "stop when every unit test in the settlement suite passes" is a verifiable goal, whereas "make settlement a bit faster" is a vague goal.

In loop engineering, the role of humans shifts from a manual craftsman to a systems architect who schedules processes.

A complete outer loop contains five operational nodes: intent definition, context collection, action, observation, and adjustment.

Loop Engineering means that instead of a human repeatedly prompting the agent to iterate, a system is designed to complete this process automatically; the human defines the purpose and completion criteria, and the AI keeps iterating until the task is done.

Loop engineering for coding agents encompasses context management, model selection, compression strategies, and termination conditions.

Loop Engineering hands the error-correction loop to agents: given a goal and checks, the agent automatically builds, runs checks, reads failures, fixes them, and repeats until all checks pass.

Simply having the model regenerate once typically yields only another version of an answer; it is only by adding a checker that one can truly converge toward the acceptance criteria.

In loop engineering, each iteration requires sending the previous results, inspection feedback, and original instructions together to the model, and the total token consumption over twelve loops can be 50–100 times that of a single call.

The loop engineering framework contains four core components: goal, acceptance, check, and budget.

Starting loop engineering practice should begin with tasks that are high-frequency, low-risk, have fixed inputs and outputs, and are easy to recover from if done incorrectly.

The check in loop engineering should be to verify results rather than to regenerate them; having the model regenerate once more often yields only another version of the answer.