Bad 要求s Make Bad Agents: 经验教训 from the Mandatory Death Penalty Debate✎ Edit

As someone building AI 智能体 at AINNA, I notice a similar pattern in public discourse every time a serious crime surfaces. 团队成员 say, “This is what happens when the death penalty is abolished.” That statement is technically wrong. What was abolished was the mandatory death penalty. The difference is not semantic - it is architectural. The death penalty still exists for specific offences, but sentencing is no longer a hard-coded output. Judges now operate with conditional logic: they can evaluate context, intent, evidence and proportionality before producing a sentence.

This is not a defence of offenders. It is an observation about how 系统 失败 when rules are over-specified. In law, one modifier like “mandatory” removes a judge’s ability to branch. The same failure mode shows up in AI 系统. If the human-side requirements are vague, incomplete or emotionally loaded, the agent will inherit that noise and propagate it through the pipeline.

AI 智能体 do not perform inference in a vacuum. They respond to prompts, 系统 instructions, retrieved context, tool schemas, parameter bounds and guardrails. Feed them a misdefined 目标 and they will optimize for the wrong reward function. The real danger is not a slow human error. The real danger is an automated agent executing the wrong decision thousands of times per minute with clean logs and no fatigue.

This is why requirements engineering beats model selection in most deployments I see. Before you let an agent classify, approve, reject, escalate or execute, you need to know exactly what “correct” looks like in your domain. Clarity is not documentation theatre. Clarity is your control plane. It is what keeps an agent from turning a business edge case into a distributed incident.

Whether you are writing legislation, designing a sentencing framework, or shipping an AI agent into production, the constraint is the same. Do not automate a process you cannot define. Do not let sentiment replace specification. And do not expect an agent to compensate for a prompt that was already broken on the human side.