A chat answers a question. A Project remembers your context so you don't re-explain it every time. An agent is a different thing entirely: you give it a goal once, and it goes and does the work in multiple steps, on its own, without you sitting there prompting each one.
What Makes Something an Agent, Not Just a Chat
A chat is a single exchange. You ask, it answers, you're both done. A Project is a standing chat with memory, it knows your team and your format, but it still only acts when you show up and type something.
An agent is neither. Give it a goal, "check this doc every week and summarize what changed," and it plans the steps itself: check the doc, compare it to last time, pull out what's different, write the summary, and hand it to you. It can use tools along the way, search, read a file, check a calendar, without you walking it through each one.
The distinction that actually matters for a manager isn't technical, it's this: a Project waits for you. An agent doesn't.
The Part That Actually Changes the Job: You Can Schedule It
This is the real shift. An agent doesn't just run when you ask, it can run on a timer. "Every Monday at 8am, check the team update doc, compile what's changed, and have a summary ready before my 9am." You never open the tool. It's just done, waiting, when you need it.
That's the leap from a tool you use to a system that runs without you, the same leap you're trying to build in your team when you delegate well: you're not doing the task, and you're not checking on it constantly either. You've set it up once, and now it happens.
Start with something narrow enough to trust quickly, not your whole job. A five-minute task you already do every week is the right size for a first one.
- Pick one narrow, recurring job, "every Monday, summarize the update doc into three bullets," not "help me manage my team."
- Open Claude and look for Tasks (or start a chat and describe what you want scheduled, Claude will offer to set it up).
- Write the instructions the way you'd brief a new hire: what to check, what to produce, and the exact format you want it in.
- Set the schedule, the day, the time, and how often it should repeat.
- Let it run once, then read the output like a first draft from a new hire, not a finished product.
- Fix what's off by editing the instructions, not by manually correcting the output every time it runs.
- Pick the same kind of narrow, recurring task, small enough that a bad first run costs you nothing.
- Open ChatGPT and go to Tasks (or ask it directly: "schedule a task that checks X every Monday and does Y").
- Describe the job, the source material it should check, and the output format you want.
- Set the schedule. ChatGPT will message you when each scheduled run completes.
- Review the first few runs closely before you trust it to just sit in the background.
- Fold anything you had to fix by hand back into the instructions, so the next run doesn't repeat the mistake.
Fill this out before you write a single instruction. If you can't fill in all six lines, the job isn't defined clearly enough to hand off yet, to an agent or to a person.
Goal: What should exist after this runs that doesn't exist now?
Source material: Where does it check, and what exactly does it read?
Output format: Bullets, a table, a short paragraph, sent where?
Schedule: How often, and at what time?
What to flag for me: What's important enough that it shouldn't wait for the next run?
What not to do: Any boundary it should never cross on its own (send externally, delete anything, make a commitment on your behalf).
Don't hand an agent your whole job on day one, the same way you wouldn't hand a new hire your whole job in week one. Start with one recurring task small enough that a bad first run costs you nothing to notice and fix. Expand what it's trusted with only after it's gotten the small thing right for a few weeks straight.
Fun fact: the failure mode with agents almost never looks like disaster, it looks like drift. A vague instruction doesn't break loudly, it just quietly produces slightly-wrong output every week until you happen to notice. Check the first few runs closely, that's when the vague spots in your instructions actually show up.