Core concepts — the two heroes

In one line. An Agent does the work; a Collection holds the work. Everything else on the platform exists to configure, connect, supervise, or display those two. You’ll be able to. Read any screen in Studio and know which of the four pillars you are looking at — and why it exists. Where this lives. Studio — the builder side of a project, opened from the Studio icon in the top bar.

Agent — the worker

An agent is a configured AI worker. Its definition is a handful of decisions, each one a tab in the agent editor:

Decision	Tab	What it means
Who is it?	Persona	Plain-language instructions: role, task, output expectations. You describe the worker; you don’t program it.
What does it think with?	Model	One of the language models registered for the project, or leave it on the project default (set under `Studio > Agents > Language Models`).
What may it read?	Knowledge	The collections attached as its grounding. This attachment is the access boundary — an agent cannot see what you don’t attach.
What can it do?	Capabilities	Tools, skills, and MCP servers. Every capability widens what the agent can act on — grant them deliberately.
What must it obey?	Governance	The guard rails attached to the agent — the rules and refusals layered onto every run.

Guard rails are written once under Studio > Agents > Guard Rails, then attached to an agent on its Governance tab.

You build one in A3 · Your first agent, test it in a live playground, and only then point it at real work. Agents come in varieties — answering, extracting, deciding, generating — covered in A4, and they compose into pipelines and teams when one worker isn’t enough.

Collection — the work

A collection is a typed home for records. You choose what it is for when you create it, and that choice is permanent — it decides how the collection behaves for the rest of its life:

Processing collections hold work in flight — each document moves through a lifecycle toward a decision.
Knowledge collections hold reference content — documents are indexed for the agent to read as soon as they land, and there is no lifecycle to walk. Large uploads take a few minutes; the collection’s Index Health tab tells you when everything is ready. This is the grounding for Agentic Search.
Structured collections hold rows — records with fields rather than pages.
Dataset (Evaluation) collections hold golden examples with the answers you expect, used when you evaluate an agent. This choice is offered only if your workspace has evaluations turned on.

Three things live on a collection and do most of the platform’s heavy lifting:

Schema — the fields you want out of every record (vendor, amount, due_date…). The schema is what agents extract into, what tables display, and what dashboards count. On this platform, the label set is the schema.
Taxonomy & tags — how records are classified and sliced.
Lifecycle — the stages a record moves through (New → Extracted → In review → Settled) and the automations that move it. Confidence gates live here: high-confidence work skips ahead; uncertain work routes to humans.

Details in D1 · Collections & schema and D3 · Taxonomy, lifecycle, tags & events.

The two surfaces

The platform has two faces, and knowing which one you’re on orients you instantly:

Studio is where builders work: collections, agents, pipelines, experiences, governance. If you can configure it, it’s in Studio.
Everything outside Studio is what your users get: the document workspace, cited chat, the pages you designed, the inbox. They never see Studio, and they never need to.

The four pillars, and where they live

The four names run across the top of Studio, after Home. Click one and its entries open in the pane on the left. Hubs and Exit Studio sit apart, on the right of the top bar.

Agents — hire and equip the workforce: the agents themselves, plus Tools, Guard Rails, MCP Servers, Skills and Language Models.
Data — shape the collections the workforce reads and fills: Collections, Ingestion, Datasheets, Views and Export.
Experience — decide what your users see and touch: General and Appearance, the Pages, Packs and Cards designer, and the Search and Chat channels.
Governance — decide who can do what and prove what happened: Users & Access, APIs, History, Snapshots, Freeze and Audit & Usage.

If one of these is missing when you look, ask your administrator — entries can be hidden for a workspace.

One concrete run, end to end

An invoice PDF lands in a Processing collection (ingestion — Data). The extraction agent reads it, fills the schema fields, and reports 0.97 confidence (Agents). The lifecycle automation sees the confidence and the matched PO and moves it straight to Settled; a low-confidence sibling routes to the review queue instead, where a person corrects one field and approves (Governance). The finance team watches both land in the document workspace and the settled-per-week dashboard tick up (Experience).

That loop — ingest, think, gate, show — is the platform. Every chapter from here is a deeper look at one part of it.

Where to go next

Build the loop yourself in S3 · Your first 30 minutes, or go straight to the pillar you need in the sidebar.