Guide
Human in the loop: what an AI agent should ask you
An agent that asks about everything isn't saving you time, and one that asks about nothing will eventually email the wrong person. Here's a simple rule for where the line goes, with the lists to give your agent.
The rule
Look up what's already been decided. Ask about what's new, risky or hard to undo.
Most questions agents ask fall in the first group: things you said on a call, wrote in a thread or told another agent. The agent asks because it can't see them, not because they're open. Give it a way to see them and most questions go away. What's left are the questions worth your time.
Look it up
An agent should find these without you, as long as it can reach a record of your work:
- Decisions already made. The price, the deadline, the format, the scope. With a source and a date.
- Your standing preferences. How you write, how you like code reviewed, what you never want done.
- Who people are. A client's role, who's on which project, who to copy.
- Where things stand. What's finished, what's blocked, what you promised whom.
- What you saw. The error in the other window, the spec you read this morning.
Ask
An agent should stop for these even when it could find something that looks like an answer:
- Money. Buying, subscribing, changing a plan, anything billed.
- Other people. Sending an email, posting in a channel, replying to a client in your name.
- Things that are hard to undo. Deleting, force-pushing, migrating data, publishing.
- Choices nobody has made. A new trade-off, a new direction, a conflict between two earlier decisions.
Drafting is fine. An agent can write the email, the post or the migration and leave it for you. The line is at sending, running and paying.
The gray zone
The hard cases are inferences: you chose monthly pricing for one client, so does that hold for the next? You approved a design last month, so is a similar one fine now? Treat an old decision on a similar question as a hint, not an answer. A good agent says “I'm assuming X because of Y; tell me if not” and keeps going, rather than either stopping or deciding quietly.
Batch the questions
One question shouldn't stall a whole job. Ask the agent to carry on with everything it can, leave placeholders for what's open, and give you the questions together at the end. Five questions answered in one minute beat five interruptions across an hour.
Write it down
Agents follow rules they can read. Put both lists in your CLAUDE.md or AGENTS.md, and give the agent somewhere to look things up. With Cachejay connected, your calls, past agent sessions, preferences and projects are searchable by any agent you use, and what you tell Claude Code or Codex is kept for next time. The setup and the exact lines to paste are in stop Claude Code and Codex asking what you already know.
The bigger picture: when your agent needs you and you're not there · context engineering for AI agents.