The honest answer to "what can I build here" is not a list of everything imaginable. It is the list of things the platform actually hands you at creation, which is shorter, more useful, and checkable in about four minutes by opening the creation flow yourself.
Here is that list.
An agent is four decisions
Every agent you build is the same four decisions in the same order.
Who it is. A name, a line on what it is for, and a description of what it does and for whom.
What it looks like. A face — picked from ready-made bases, or generated from the name — and a presence tier that decides how alive that face is.
How it speaks. A voice, bound to the agent at creation, and an instruction body: compiled for you from what you already typed, started from a proven template, or written yourself.
What it can do. A set of capabilities, ticked from a grouped list.
That last one is the part people underestimate, so it is worth reading properly.
The capabilities, as they are grouped today
Knowledge and memory. Web search, for current information. Memory recall and memory store — the two halves of an agent that remembers you between conversations rather than starting fresh every time. Consult brain, which sends the agent to its own knowledge before it answers. Wiki lookup and code wiki, for agents wired to a documentation source.
Three of these are ticked for you by default: web search, memory recall, memory store. That default is the platform's opinion about what a useful agent is — one that can find out, and one that remembers.
Communication. Send email on your behalf. Make calls. Check calendar availability.
Creation. Generate images on request. Creative media and social posting, both managed from the Tools panel.
Commerce and wallet. Read a wallet balance, spend within limits, confirm a pending spend, read the history and the spending limits. List products and prices from a catalogue, create a checkout link, check a payment's status, report configuration.
That last group is marked Advanced and every one of its members is off by default, with a note saying commerce and wallet capabilities must be enabled per agent deliberately. An agent that can spend is a decision somebody makes on purpose, not something they inherit from a default.
So what does that add up to
Read the groups back and the shape of what is buildable is clear enough.
An agent that answers a phone or a front desk. Communication plus memory: it can take the call, know who called last time, and put something in a calendar.
An agent that knows one body of material cold. Consult brain plus wiki lookup, wearing a face and a voice, sitting on the page where people get stuck.
An agent that produces work. Creation capabilities, for the drafting jobs that eat afternoons.
An agent that transacts. The commerce group, on purpose, with limits set on purpose.
An agent that is simply good company. A face, a voice, an instruction body, and memory. Nothing in that list is small.
The parts you do not have to build
Two things arrive free with every agent, and they are the two most people expect to fight for.
A place to live. Publishing gives your agent its own page, a share link, and an embed widget for a site you already run — the same three doors, switched on together by one button.
A presence. The face, the movement, the voice. You do not commission an avatar or wire up a renderer. You pick, and the platform paints it.
Where the honest edges are
Three, and knowing them beforehand is worth more than any feature list.
Two of the six presence tiers are studio-authored. ASCII and Sketch Line are made in the studios after the agent exists. Their cards in the creation flow are visible but disabled, and they say so themselves. Kinetic Typography once sat with them, but it is create-time now — the agent's words are its face, so no authored art gates it, and the tier walk-through covers what that looks like.
The face is chosen once. It and the presence tier are written together as the agent's presence identity at creation, and there is no owner-facing screen today for swapping them afterwards. Choose deliberately.
Some capabilities are managed elsewhere. Creative media and social posting are ticked in the creation flow but configured in the Tools panel — ticking is intent, not wiring.
Start smaller than you think
The most common mistake is building the agent you eventually want instead of the one you can judge tomorrow.
Give it one job with an obvious right answer — the question your customers ask most, the task that eats a specific hour of your week. Ship that, watch it, and let what you learn choose the second one. The creation flow takes minutes, versions only ever grow, and nothing you make is overwritten by what you make next.