Agent sidebar vs Tab — two ways in

Cursor is installed, signed in, and reading budgetcli. Before you point it at the date bug, spend two minutes on a distinction that quietly decides how productive the next hour is: the Agent sidebar and Tab are not the same tool, and they’re not good at the same things. Most early frustration with Cursor comes from using one where the other belongs. You met both in passing on the chapter’s opening page; here we make the split sharp enough to act on.

Tab: the autocomplete that finishes your thought

Tab is the grey ghost text. You’re typing, and Cursor offers the rest of the line — or the rest of the block — which you accept with the Tab key or ignore by typing past it. It’s reading the file around your cursor and the shape of what you’ve started, and it’s fast: no prompt, no waiting, no conversation.

What Tab is good at is continuation. You’ve written a function signature and it drafts the body. You’ve started a test and it fills in the assertions. You’re halfway through a repetitive block and it finishes the pattern. You stay in the driver’s seat, typing, and Cursor keeps pace half a step ahead. It’s the descendant of the ghost-text completion you already had in VS Code — recognisable the instant you start typing, just sharper at predicting where the next edit goes, not only the next character.

What Tab is not good at is anything you’d have to explain. It can’t take an instruction. It can’t reason across files you don’t have open. It can’t be told “do it the way the rest of this repo does it.” If you find yourself wishing you could tell the suggestion what you want — that’s the signal you’ve reached the edge of what Tab does, and the Agent is the tool you actually want.

Tab earns a chapter of its own. The full treatment — how to read its predictions, when to lean on it and when to ignore it, and how it reshapes your typing rhythm once it’s tuned — lands later, in the daily edit loop. The daily edit loop chapter is where Tab stops being a curiosity and becomes muscle memory; here you only need to recognise it as the continuation surface.

The Agent sidebar: the colleague you brief

The Agent sidebar is a conversation. You open it with Cmd/Ctrl+I, describe what you want in words, and Cursor proposes — an explanation, a diff, a multi-step change — which you review. It reads far beyond your cursor: the open files, files it decides are relevant, and (in later chapters) the rules you’ve written and the tools you’ve connected.

The Agent is the surface this whole course is about, because it’s where Cursor stops being autocomplete and becomes an agent — something you delegate to, not just type alongside. It doesn’t answer from the prompt alone; it goes and reads first, then proposes, then waits for you before it touches your disk. That read-propose-review rhythm is the engine under every task you’ll give it, and it’s worth watching once in slow motion before you trust it with a real change — which is exactly the next lesson. Watch one turn of the loop takes the Agent apart beat by beat; for now, the only thing to internalize is the split:

• Reaching for a continuation of what you’re typing? Tab. Stay in the file, let the ghost text help, accept with Tab.
• Have something to explain — a bug to describe, a change to request, a question about unfamiliar code? Agent. Open the sidebar with Cmd/Ctrl+I and brief it like you’d brief a teammate.

Why this matters on inherited code

Here’s where the distinction bites for your situation specifically. budgetcli is code you didn’t write — a friend’s weekend budgeting API now running your own money. The fastest way to get oriented in an unfamiliar repo isn’t to start typing and hope Tab catches your drift — it’s to ask. The Agent can read the code and explain it back to you, trace where a value comes from, or summarize what the CSV importer does, before you change a single line. Tab can’t do any of that, because there’s nothing to continue yet — you don’t know enough to start the line.

That’s exactly the situation in front of you: the import path mangles some dates — a statement that clearly says one day lands in the database as another — a bug you can describe but couldn’t confidently fix by typing. So this one’s an Agent job. The Tab key stays useful all week for the keystrokes you were already going to make; the date bug is work you’d rather describe than type, and that’s the Agent’s job.

You now know which surface fits which moment. Next, open the Agent sidebar and watch one turn of its loop close — read, propose, review — on a harmless question, so the real fix that follows holds no surprises. Next: understand one turn of the loop.