What the Agent Actually Is

It has tools, and it uses them

The agent has a large registry of tools — real operations against your workspace and your connected accounts. Reading tasks. Creating them. Searching mail. Reading the calendar. Running reports. Writing to connected third-party tools. Hundreds of distinct capabilities, organized into groups.

When you ask it something, it does not answer from memory of your workspace. It goes and looks, with a tool call, at the moment you ask. That is why it can tell you what is overdue right now rather than what was true when someone last pasted a summary in.

And when you ask it to do something, it does it by calling the same underlying operations the interface calls when you click. There is no separate agent-shaped back door: it is a user of the product.

It is inside the boundary, not outside it

The agent operates on the ACTIVE workspace, and it sees what that workspace contains.

This has an immediate practical consequence, which is that the answer to 'why can't it find my task' is very often 'you are in a different workspace'. Check the switcher before anything else.

It also means the agent is subject to structure rather than exempt from it. Per-project access, the charter, the action policies — these are not advisory prompts asking it nicely. They are enforced at the point the tool runs. A charter Negation refuses task creation regardless of how the request is phrased, because the gate is on the capability and not on the wording.

Setup quality is agent quality

This is the sentence worth taking away from the lesson.

An agent that can see a workspace with real projects, real assignees, real due dates, a connected calendar and a written charter can answer questions about your business. An agent that can see one project called 'Stuff' with twelve untitled tasks can answer questions about general project management.

Same agent. The difference is entirely what there is to look at.

That is why Modules 1 through 6 come before this one, and it is why the most common cause of 'the AI isn't very useful' is a workspace with nothing in it. The fix is never a better prompt.

Everything it does is recorded

Agent actions go through a single chokepoint, and that is where the ledger is written. Every action the agent takes is logged: what it did, when, with what arguments, and what happened.

That matters for three reasons. You can audit it — 'what did it actually do' is a question with an answer. You can undo much of it, because the same actions go onto the server-backed undo stack. And the record is what makes it reasonable to let it act at all: an agent whose actions were invisible would be one you could never safely widen.

Lesson 7.4 covers reading that record. For now the point is that it exists, and that you never have to reconstruct what happened from memory.

What it is not

It is not a replacement for deciding. It will draft, propose, summarize, schedule and execute — but the four questions in triage, the charter, the priorities and the trade-offs are yours. The product is built on that split deliberately.

It is not omniscient about things outside the workspace. It has web search and can read pages, but a fact that lives only in someone's head or in a system you have not connected is a fact it does not have.

And it is not silently autonomous. What it may do without asking is governed, defaults conservative, and is entirely under your control. That is Module 8.