Where an AI agent should ask for human approval
CodeAstra Tech · Engineering decision guide
Start with the problem
Distinguish suggestions from actions. Drafting a response and sending it externally carry different consequences, even when the text is the same.
Define the scope
Define allowed tools, permitted records and action limits. Require approval for consequential external writes, financial actions or irreversible changes unless a bounded policy explicitly allows them.
Check the failure cases
Show the proposed action and its exact target before approval. Log the decision, prevent duplicate execution and stop when a tool response is ambiguous.
Make the decision
Test rejected requests, repeated calls and unavailable tools. A demonstration of the successful path is not evidence that the agent is safe to run unattended.