HACH developer: gate AI agent actions by reversibility, not blanket prompts

The design starts from a problem the developer says is unavoidable once a chat-resident agent is given tool-execution rights: asking about everything wears users down, so they click yes without reading, and the prompt stops working as a safety device. The answer is not to cut confirmations but to make them scarce and meaningful, by sorting actions into freely_reversible, reversible_with_cost and irreversible, and by judging reversibility by what the user sees rather than by what the API allows.

The second axis is blast radius, whether an action touches only the user, the team, or the outside world. Combined with reversibility, that yields a matrix that fixes confirmation in advance, so adding a new tool means setting two classifications rather than debating each case. When a prompt is needed, it should carry the detail that only becomes visible at the moment of commitment, such as the recipient address or the target channel.

To keep repeated work from sliding back into fatigue, the fourth pattern records an approved action-plus-target pair as a trusted pattern and skips or lightens confirmation next time, reverting to a full prompt if the target or its visibility changes. The developer frames this as complementary to limiting which tools an agent holds at all: permission design decides what the agent can do, while confirmation design decides when a human must step in just before an action becomes irreversible.