Three things your AI agent still breaks on.

The exact clicks

1. Find where your agent's starting instructions live (its "system prompt" — the text it reads before your first message). In ChatGPT: Settings → Personalization → Custom instructions. In n8n or Make: the agent node's own "System Message" field.
2. Add: "At the start of each session I'll tell you what changed since last time — don't assume you remember on your own."
3. Add a second line: "If a site asks you to log in, stop and ask me — never guess a password or click through a login screen yourself."
4. Add a third line: "Always create a draft. Never send anything yourself — I'll click send."
5. Save, then test each one on purpose: point it at a login-walled page and give it something to send, and confirm it stops both times instead of acting.

What changes

• It forgets everything between sessions by default — the memory line fixes what it's told, not what it retains.
• It gets stuck, or worse, guesses, the moment a site asks for a login — the instruction makes it stop instead.
• It can send something you never meant to send, with no undo — the draft-only line puts the final click back on you.

What it will not do

• These are three instructions, not three features — the agent still has no real memory, no login handling, and no undo. The lines only change what it's told to do about each.
• None of this needs a rebuild, but it does need you to actually add the lines — an agent left on its defaults still has all three problems.