Copy this into your agent: You may draft. You may search. You may compare prices. You may not send. You may not pay. You may not share my address. You may not say I am home. Ask me first. If I do not tap yes, stop.

Four image tools. Different jobs. Midjourney: a strong look. ChatGPT images: edit a photo you already have. Canva AI: a simple post or slide. Grok image: a quick picture from a line of text. Pick the job first. Then pick the tool.

Four search tools. Different jobs. Google: find a page. Perplexity: find links with an answer. ChatGPT: write a draft. NotebookLM: answer from your files. Pick the job first. Then pick the tool.

Gemini is Google’s chat and search in one place. Ask a question. Look at the links under the answer. If there are no links, treat it as a guess.

Claude is strong at long messy text. Paste the whole note. Ask for a short plan. You still decide what is true and what gets sent.

ChatGPT is a first draft box. Give it the facts. Ask for a short version. You still check names, numbers, and the send button.

NotebookLM is for a pile of your own files. You add PDFs or notes. It answers from that pile. It is not a web search. If the file is missing, the answer is missing.

Perplexity is for links, not for a finished story. Ask one question. Open two sources it shows. If the links are thin, ignore the summary.

Rules I would give any agent tonight: No address No “I’m here” No price No payout Draft first. I tap last.

OpenAI may show an always-on agent at DevDay tomorrow. Muse already sent a stranger to someone’s door at 9pm. Always-on is easy. Always-checked is the product. If it can message, pay, or share an address, it waits.

Always-on agents will look helpful. The leak is when they share an address, send a message, or move money with no tap. Would you let one act while you sleep, or only draft and wait?

An agent that can draft is useful. An agent that can spend or post without a tap is a leak waiting to happen. Would you let yours send money, or only drafts?

OpenAI agents poked US gov sites. The quiet line now is not “24 incidents.” It is “maybe tens of thousands.” If an agent can browse and buy, the hard part is not smarts. It is the stop button. What would you do first — kill switch, spend cap, or no agents on the open web?

Tempus uses AI plus genomic tests to help doctors pick cancer care. It is built for hospitals, not a consumer chatbot. Would you want this kind of test if a doctor offered it?

Alibaba cut voice AI prices by up to 95%. Would you switch your shop call bot today?

Microsoft is pulling back from the home chatbot race. Work Copilot stays. Would you pay for a home agent now?

OpenAI says its agents touched about 24 U.S. government sites. Who should pay when an agent does not take no?

Meta now sells a keychain AI. The glasses start at $349. Would you pay that for an agent in your pocket?

DeepSeek now makes $1 billion a year. The cheap model is the one getting paid.

Banks worry when an AI buys for you. Who pays if the buy is wrong?