What is this article talking about?
AI is increasingly able to work continuously and automatically: after completing one step, it takes the next step, assigns its own sub-tasks, and checks and accepts itself. The exciting part is speed, but the scary part is also fast: a wrong judgment that has not been challenged will be amplified by every subsequent lap. This article compiles five loop guardrails that I actually use. TheyControl Engineering’s boundary thinking andLoop Project’s improvement thinking is connected together. Who is
· People who have asked AI to automatically run consecutive tasks and are beginning to worry that it is "too smooth and suspicious"
· People who often hand over work to the second AI or sub-agent are found to often run the clock after handing over.
· Supervisor who leads the team to use AI: These five points are also acceptance language for human teams
· Five guardrails that can be copied directly into your own rule base
·Each sentence is attached with a usable sentence structure or example, you can use it just by pasting it.
· A habit of judgment: when you see a popular AI framework, how can you absorb the concepts without moving the entire set?
List of five guardrails
Major conclusions that have not been challenged by the second perspective can only be marked as hypotheses, not facts.
The three perspectives of logic, failure, and over-engineering take turns to attack their own conclusions.
Non-goals Specifies what not to do, allowing the path to draw where it can touch.
If the same method fails twice, change the route and do not do a third blind trial.
"Looks okay" is not evidence. Before calling it a day, ask: What level of evidence is it?
Do small things like checking the time and changing the format directly; only put guardrails on major conclusions, handovers, and high-risk operations.
Guardrail 1: Undefended presumption, mark first and then discuss
Major conclusions given by AI (root causes of errors, architectural solutions, business suggestions, security judgments) can only be marked as "undefended assumptions" before being challenged by a second perspective. It may be all correct, but "has not been challenged" is important information in itself. Written as a confirmed fact, every subsequent step will build a house on the wrong foundation.
This guardrail has the lowest cost and the greatest effect: it allows the person (or AI) who takes over later to know at a glance which words can be stepped on and which words should be ignored a priori.
Guardrail 2: Defense from three perspectives, attack yourself in turn
The defense uses three perspectives to look at the same conclusion in turn. Each perspective has its own list of questions:
- Logical perspective (skeptic): Is this conclusion based on a single piece of evidence? Have you ever regarded correlation as causation? Has contrary evidence been ignored? Do acceptance conditions really prove the conclusion?
- Failure perspective (red-team): What’s the worst case scenario? Will you encounter irreversible things such as permissions, keys, data loss, public release, and money? How to restore if it fails? Are there any shortcuts that may seem feasible but magnify the risks?
- Over-engineering perspective (simplifier): Is there a smaller first step? Is the one-time task made into a permanent system? Can it be done first using existing tools and manual acceptance?
An important mark of honesty: letting the same AI play three perspectives to attack itself is called "self-attack only", which is better than nothing; the real defense is to submit it to another company for model review. Sub-agents of the same model still share the same set of blind spots when reviewing each other.
Guardrail Three: Two columns are added to the handover package to prevent the task from being easily expanded.
When handing over work to the second AI, subagent, or the next round of automated tasks, the handover instructions often describe what to do, but not what not to do. The executor "doing more easily" sounds hard-working, which is the main source of accidents in the automatic loop. Therefore, there are two columns for each handover:
I use these two columns in every assignment I make. Based on my own experience, the effect is very direct: the executor's creativity is guided to "do the specified thing well", and when it exceeds the scope, he will stop and ask questions, and the number of clock ticks is significantly reduced.
Guardrail 4: Change road and brake, if failed twice, turn
AI has a bad habit much like humans: if the same method fails, try again, fail again, try again. No one interrupts it in the automatic loop, and it can blindly test all night long. So digitize the brakes:
- The same method failed 2 times: change roads, upgrade tools, or stop and report back, without doing the third blind test on the same road.
- Sub-agent returned substantive error 1 time: The main control withdrew the judgment and did not allow the same sub-agent to blindly run the second round.
- If "should be possible", "seems okay" or "maybe" appears but no evidence can be provided: it will be regarded as unverified and cannot be regarded as completed.
Numbers are important. Writing "If you get stuck, change the method" is equivalent to not writing it, because the AI (and humans) will always identify "stuck" later than it actually is.
Guardrail 5: Grading of acceptance evidence, ask a question before calling it a day.
The last step of the cycle is acceptance, and the most common falsification (unintentional) of acceptance is "I checked it." Therefore, the acceptance report must indicate which level of evidence it is, from strongest to weakest:
- First level machine acceptance: test passed, script output, file existence, counting results, deployment status.
- Second level data re-examination: reading files section by section, comparing sources, and comparing counts.
- Level 3 Cross-family review: Opinions of another company’s independent review of the model.
- Level 4: Manual visual inspection or inference: can only be used as low-strength evidence and cannot alone support high-risk "completion".
How you can start: Put guardrails on your next big task
There is no need to hit all five at once. Pick your next important task (one that needs to be handed over, one that will involve formal information, one that the conclusion will be used for a long time), and install it in this order:
- Before starting construction: Complete the handover instructions with the columns Non-goals and Allowed Paths.
- In the process: Make an appointment with AI to change roads and brake. If the same method fails twice, you must stop and report back.
- When drawing a conclusion: It is required to mark the "undefended hypothesis" and then ask it from three perspectives.
- Before finishing the work: Ask "What level is your evidence?" The fourth level of evidence will be added to the next level.
After running smoothly, write these five items into your rule library or skill package, so that every cycle will automatically run with guardrails. How to recover when an accident occurs in the loop, another articleLoop Safety and RecoveryYou can continue reading.