Notes on building an agent that acts for someone

RLS protects tables, not functions. Our SECURITY DEFINER credit RPCs were callable with the anon key. What Supabase's Oct 30 change leaves open, and our fix.

Our first cold email batch was accepted by SMTP and reached 0 inboxes. The 4 DNS findings behind our DMARC fail, and how we checked each one.

Most agentic marketing writing assumes a team for the agent to plug into. What changes when the agent is the team, and the only reviewer has the least time.

Domain authority takes years and AI answer engines do not wait for it. What a new domain can do to get cited, and what it should stop pretending will work.

The backlash is not about AI writing. It is about publishing without judgement. The distinction matters, and it is a process problem before a model problem.

You describe what you want and the code appears. Marketing resists the same treatment for a specific reason, and it is not that the models are worse at it.

Article 50(2) requires generated media to be machine-readable as artificially generated. What that costs in code when your encoder gives you no metadata hook.

The approval boundary is not a missing feature — it is the load-bearing design decision in an agent that acts for you. What it costs, and what it buys.

Seven failure modes in autonomous agent platforms: phantom completions, credit burn, unauthorised action, credential concentration — and the fix for each.

Per-seat pricing breaks when software does the work. Where to put the charge, why the attempt is the wrong place, and the idempotency problem nobody names.