AI operations agent with permission-gated actions Problem: Businesses want AI agents that take real actions — sending emails, updating records — but handing an LLM unsupervised write access is a legitimate trust problem most teams correctly won't accept.

Approach: Built a governed agent for a hospitality studio use case, with a decision layer that checks whether a request is a new action or a question about something that already happened before touching any tool — the failure mode that caused it to draft a new email when asked whether one had already been sent. Sensitive actions hold for human approval, every action is logged in a full trace, and the agent verifies an action actually succeeded before reporting it did.

Outcome: A working agent that does the reasoning while a human keeps the final say on anything that matters — nothing sent or changed without a trace, and no false "sent successfully" confirmations.

Tools: Next.js, Groq, Supabase, pgvector