A like is not a relationship: where our agent's permission to reply stops

The developers behind a public-facing AI agent have shared their approach to managing automated replies, emphasizing that 'warm' interactions do not automatically grant permission for unsupervised communication. The team implemented a strict approval gate where an AI drafts replies, but a human must approve them unless specific criteria are met. By distinguishing between 'warm' relationships (based on engagement ledgers like likes and quotes) and 'exempt' status (reserved for established connections), the team ensures that the agent remains honest about its human-in-the-loop nature. The system uses a robust audit trail, hashing draft bodies to prevent unauthorized changes. This design prioritizes safety and transparency, enforcing strict boundaries at the send path rather than the drafting stage. The authors argue that developers should treat permission as a data-driven constraint, ensuring that every widening of automated capabilities is backed by an explicit, dated human decision.