←back to thread

419 points serjester | 1 comments | | HN request time: 0.209s | source
1. cadamsdotcom ◴[] No.43540458[source]
Models aren’t great at deciding whether an action is irreversible - and thus whether to stop to ask for input/advice/approval. Hence agentic systems usually are given a policy to follow.

Perhaps the question “is this irreversible?” should be delegated to a separate model invocation.

There could be a future in which agentic systems are a tree of model and tool invocations, maybe with a shared scratchpad.