Assess the risk of the task and come up with protection against agent errors. Return JSON {"task":"...","riskLevel":"...","riskyPoint":"...","safeguard":"..."}: task - task, riskLevel - one of: low, medium, high, riskyPoint - the most dangerous place (where an invention or irreversible action is likely), safeguard - how do you check it.
Agents have predictable weaknesses, and knowing them is half the battle. First: a confident tone when making up your mind. If there are not enough facts, the agent will still give a smooth answer - just a made-up one. Second: looping - he can walk in circles, repeating one step if the target is blurred. Third: a wrong assumption at the beginning, which causes all the work to go in the wrong direction. The defense mechanics are based on checking. Insider trick: require a link to the source for each important statement and a direct note of uncertainty - “if you’re not sure, write so.” It is easier to catch a fiction when there is a link nearby. The second technique is a budget of steps and a separate verification step: “do it, then double-check the result according to this criterion.” Limiting steps prevents you from going into an endless loop. A common mistake is to trust an answer because of its confident appearance and convincing details. Persuasiveness and truth are two different things. Assess the risk of the task in advance: the higher it is, the stricter the verification and the less autonomy you should give the agent.
Unlock access to submit solutions for instant AI review.