Agents

An agent is something that pursues a goal on a schedule, inside a budget you set, and reports back with its sources.

You describe what you want kept an eye on. You get something that does it every week, tells you what changed, and shows you where each claim came from. It costs at most what you agreed to, and it cannot publish anything or send anything to anyone.

What makes it an agent

An agent decides for itself which sources to look at and when to stop. That is the whole difference from a fixed workflow. A workflow does the same five things in the same order every time; an agent is given a question and a budget and works out the route.

That freedom is why the rest of this section exists. An agent that chooses its own path needs a hard envelope around it, a record of what it actually did, and a person who says yes before anything it produced can be used.

	Agent	Playbook
You define	a goal and a budget	the exact steps
It decides	which sources, how many, when to stop	nothing
Output	a note, with sources	whatever the last step returns
Best for	research, watching, finding, digesting	repeatable known work

If you can write down the steps, write down the steps. A playbook is cheaper, faster and completely predictable. Reach for an agent when the interesting part is the judgment.

The four rules

Everything in this section follows from four properties that hold for every agent, with no exceptions and no per-agent setting to turn them off.

Created switched off. A new agent never runs until you turn it on. Creating one costs nothing. You can run it once by hand, read what it produced, and only then put it on a schedule.

A hard ceiling per run and per month. You see both numbers before you agree to anything. The monthly one is enforced: when it is reached the agent stops itself and tells you, rather than continuing and invoicing you for it.

Nothing goes outward. No agent sends an email, posts anything, writes to your CRM, or publishes a page. Every result is inward: a note you read. See Approvals and veto.

Every claim carries a source, or it is held back. An agent that cannot point at where a figure came from does not get to state it. That output is quarantined and waits for you. See Approvals and veto.

Starting points

Page	What it covers
Your first agent	Create one, run it once, read what it produced
Archetypes	The six shapes, what each is for, what each costs
Limits and costs	The envelope, the monthly ceiling, what happens at the ceiling
What agents can see	The data statement: what an agent may read, and what it may never reach
Approvals and veto	Inward versus outward, and how to reject something
Testing an agent	Run once, read the narrative, decide

Where you use them

In Studio under Agents, or from any AI assistant connected over MCP, where the same agents appear as tools you can call by name. Both surfaces reach the same agent and the same limits. There is no separate configuration.