Blog
When your AI agent needs you and you're not there
Agents now work for an hour while you're in a meeting, at lunch or asleep. Most of them still stop at the first question only you can answer. Most of those questions you've already answered somewhere: on a call, in a thread, in another agent's session.
Where agents get stuck
You give Claude Code or Codex a task and step away. Twenty minutes in, it hits a fork: monthly or yearly pricing on the new plan? Does the client want the logo on the left? Who signs off on the copy? It either stops and waits for you, or it guesses.
Waiting wastes the hour you stepped away for. Guessing means redoing the work. And the answer usually exists. It was said on Tuesday's call, or decided in a Slack thread you read, or settled in last week's session with a different agent. The agent just can't see it.
The four things agents ask you for
Watch what agents interrupt you with and it comes down to four kinds of question. Each is something Cachejay already keeps as you work:
- What was decided. “Did we agree on the yearly price?” Cachejay has your calls, transcribed with no bot, and the text of what you read and wrote. The agent searches it and gets the quote and the date.
- How you like it done. “Tabs or spaces? Staging first?” What you’ve told your agents across projects is merged into one profile, so a rule you gave Claude Code last month reaches Codex today.
- Who someone is. “Who’s Dana, and should she be on this email?” Cachejay keeps the people in your work: who they are, which projects they're on, what's open with them.
- Where things stand. “Is the webhook fix done?” Each project has a page with its status, recent agent sessions, calls and open to-dos.
What it looks like
A made-up example. You asked an agent to build the pricing page for a bakery's ordering app, then went into a meeting:
› build the pricing page from the plan we agreed ⎿ cachejay · search_memory Found it: on Tuesday's call with Northwind, you agreed $99 a year, starting in November, with SSO on the team plan. Building that. One thing I couldn't find: whether the team plan has a seat minimum. I've left it as a placeholder and noted it below for you.
Two things to notice. The agent said where the answer came from, so you can check it. And it didn't invent the one thing nobody had decided. It kept going and left you a question for when you're back.
What still comes to you
Cachejay doesn't decide anything. It answers with what you already said, saw or agreed. Some things should still wait for you, even when an answer could be found:
- Anything that spends money or signs you up for something.
- Anything sent to another person in your name.
- Deleting things, or anything else hard to undo.
- A choice nobody has made yet. An old decision on a similar question is a hint, not an answer.
The goal is fewer interruptions, not an agent that acts as you. We wrote up the line between the two in what an AI agent should ask you.
Answers that stick
It works in the other direction too. When an agent finishes a piece of work, it tells Cachejay what's done, what was decided and what's next, and the to-dos it finished get ticked. The project's status is rewritten from that. When you answer a question in Claude Code or Codex, that session is kept as well, so the next agent, in any app, can find your answer instead of asking again.
Setting it up
It takes a few minutes. Install Cachejay, choose Connect AI Apps in the menu bar, and pick your agents. The step-by-step version, with the lines to add to your CLAUDE.md or AGENTS.md and three questions to test it with, is in stop Claude Code and Codex asking what you already know.
More on the idea: the human cache · agent context. Or try Cachejay, 14 days free with no card.