Notes on working with AI
Agents in the Claude chat: from asking questions to handing over jobs
The Claude chat can take on a whole job. It works through the steps, picks the tools and hands back finished files. Here's how to brief it and stay in control.
Four rungs down a vertical spine: a question, then a whole task, then a doc or a deck, and finally a scheduled task. Beside each rung is what you no longer do yourself. With a question, nothing: the answer comes back and the rest is yours. With a task, Claude does the steps. With a doc or a deck, the output is kept too, at one link. With a scheduled task, the request is saved too, so you stop asking. A circular arrow at the end marks the last rung running again without being asked.
What you stop doing yourself
A question
Nothing. The answer comes back, and every step after it is yours.
A whole task
The steps. Claude searches, reads the files, runs code and hands back a finished file.
A doc or a deck
Keeping the output. In Docs or Slides it stays at one link you edit together, instead of scrolling away in the chat.
A scheduled task
Asking. The request is saved, and it runs again in the cloud on the schedule you set.
Runs again, no asking twice
That's the difference between asking a question and handing over a job.
The Claude chat does more than answer questions now. There's no mode to pick. You ask for what you need in any conversation, and Claude decides whether it's a quick answer or a task. When it's a task, Claude works through the steps on its own. It searches the web, reads your files, runs code, and hands back finished files: documents, spreadsheets with working formulas, decks that open in PowerPoint.
Inside a task, Claude picks the tool for each step. A connected app comes first, then a browser, and working directly on your screen only as a last resort. Deep questions go to Research, a built-in agent. Type /deep-research, or pick + and then Research, and you get a cited answer in minutes. When a step needs a website, Claude opens its own browser in a side panel, and you can watch it click through and fill in forms. It doesn't see your own tabs or passwords. Your files, the browser and computer use need Claude Desktop open. Everything else works on the web and on mobile too.
When the result is a document or a deck, it lives in Claude Docs or Claude Slides. You ask for one, edit it together, share it by link, and export it to Word, PowerPoint or PDF. Both are in beta, on paid plans.
You can have several tasks running at the same time, each one its own conversation in Recents. On a big task, Claude can split the work and run several sub-agents at once. You see that as a plan and step-by-step progress, not as a stack of separate windows to manage. You don't have to sit and watch, either. Big tasks keep running in the cloud, and from your phone you can check in, answer Claude's questions or redirect it.
A quick question still fits when you need one piece, like a fact or a better way to phrase a paragraph. Hand over the whole task when what you want at the end is a file and getting there takes several steps across a few sources. Bigger tasks do use more of your plan than quick questions, so group related work into one task rather than spreading it across five.
Brief it the way you'd brief a colleague. Hand over the outcome, not the steps. Say what you want to end up with, the format, and what Claude should work from: the files, the links, the apps. 'Summarize the market' is a question. 'A two-page brief on how our three main competitors price, as a doc, built from these five links and the sales file from last quarter' is a job Claude can finish.
You decide how often Claude checks with you. Manual is the default, and it asks before each action. Auto keeps going and runs safety checks along the way. Start on Manual, and switch to Auto only for work you'd be fine undoing. When a task sends messages, spends money or changes important files, keep a close eye on it. Claude asks before permanently deleting files, and you can stop it or change its direction at any point.
Work that repeats becomes a scheduled task. In any conversation, write something like 'every Monday at 9, update the pipeline report', confirm, and click Schedule. It runs in the cloud and lives under Scheduled in the sidebar. Anthropic's own example covers the whole arc in one conversation: a pipeline report plus five matching slides, then a schedule that runs it every Monday. You brief it once. After that, the report just shows up.
Frequently asked questions
What are agents in the Claude chat?
It's Claude taking on a whole task inside an ordinary conversation, instead of answering one question. You ask for what you need, and Claude decides whether that's a quick answer or a task. On a task it works through the steps on its own, picks tools like Research, its own browser or a connected app, and hands back finished files: documents, spreadsheets with working formulas, decks that open in PowerPoint.
Can Claude work on several tasks at the same time?
Yes. You can have several tasks running at once, and each one is its own conversation in Recents. On a big task, Claude can also split the work and run several sub-agents in parallel. It all happens inside that task's conversation, where you can see what Claude is doing at each step.
How do I get Claude to run a task every week?
Make it a scheduled task. In any conversation, write something like 'every Monday at 9, update the pipeline report', confirm, and click Schedule. It runs in the cloud and lives under Scheduled in the sidebar.
How do I stay in control while Claude works on its own?
You set how much it does without asking. Manual, the default, asks before each action. Auto keeps going with safety checks. Start on Manual, move to Auto only for work you'd be fine undoing, and keep a close eye on anything that sends messages, spends money or changes important files. Claude asks before permanently deleting files, and you can stop or redirect it at any time.
Keep learning
Agentic workflows in the Claude chat
From asking questions to handing over jobs.
View topic