Agents
An agent is the assistant you chat with. Where a model is the raw LLM that answers a request, an agent is that model wrapped with instructions and settings — a system prompt that tells it how to behave, a provider and model to run on, and a handful of behavior knobs. Two agents can run on the very same model and still act nothing alike, because the agent is where the personality and rules live.
Put simply:
- A model is what answers — the LLM, with its context window, capabilities, and cost.
- An agent is who answers — that model plus a system prompt, a name, and its behavior settings.
An agent sits on top of the provider → model chain — it picks up a model, which belongs to a provider — so the natural setup order is connect a provider → add a model → configure an agent that uses them.
Two tiers: system and workspace
Agents mirror the provider and model tiers. Brisal ships a small built-in catalog of system agents, and you clone one into a workspace to get an editable copy of your own.
System agents are built into the app. They ship with the binary, they’re read-only, and they act as starting points — you can’t edit or delete them. Today the catalog holds two built-ins: Le Chaton Fat, a sharp, complete general assistant and the app’s default, and Le Lazy Chaton, the same underlying assistant tuned for terse output. System agents wear a Shipped badge.
Workspace agents are yours. Each lives inside a single workspace, and every field is editable — the name, the system prompt, the model it runs on, its behavior settings. You get one either by cloning a system agent or by creating a custom agent from scratch.
One mechanic, three verbs: you connect a provider, add a model, and clone an agent. All three copy a built-in into your workspace so you own the copy — only the word changes with the thing.
Cloning an agent into a workspace
You manage a workspace’s agents from its Agents tab. At the top sits an Available agents catalog — “Clone one to your workspace to customize, or create a custom one” — with the Shipped system agents (each with a Clone button) alongside a Custom tile (Create). Below it, a Default agent line and your own workspace agents.
The workspace Agents tab: the Shipped catalog on top, your workspace agents below. A fresh workspace has none of its own yet.
There are two ways in: Clone (or Create) straight from a catalog card, or the New Agent button, which opens a chooser first:
- Clone from system catalog — copy a bundled agent (e.g. Le Chaton Fat) into this workspace.
- Create custom agent — configure a new agent from scratch (see below).
Cloning shows the source agent’s summary (provider, model, thinking, tags) and
asks for a new id for your copy (a short slug, defaulting to the source id
with a -copy suffix; it must be unique in the workspace and match
^[a-z0-9-]{3,32}$). Confirm, and the copy lands in your Agents tab as an
ordinary, fully editable workspace agent.
Clone-then-own: what you get
Cloning an agent is a copy, not a link — the same clone-then-own mechanic as connecting a provider or adding a model: you own every field (name, system prompt, model, behavior settings), and there’s no auto-update if the built-in later changes. The copy keeps no reference back to its source — from the moment you clone, the two evolve separately.
The agent’s instructions
The system prompt is the heart of an agent — the standing instructions it carries into every conversation. This is where you set its role, its tone, the rules it should follow, and anything it should always keep in mind (“You are a senior Rust reviewer. Be terse. Prefer the project’s existing conventions.”).
You set it in the agent form’s Behavior section. It’s optional (an agent with no prompt falls back to the model’s default behavior), supports Markdown, and holds up to 64 000 characters — long enough for detailed playbooks, examples, and formatting rules.
Choosing a provider and model
An agent’s Provider & Model section is where it binds to the LLM it runs on. Pick a provider first; the model dropdown then lists the models available on that provider — the workspace models you’ve added to it. If the dropdown reads No models available, that provider has none yet, so go add one.
Both the provider and the model are optional on the agent — you can leave them unset and choose at the moment you start a chat. But an agent with a fixed provider and model is the common case: it means “this assistant always runs on this LLM.”
Thinking and behavior
Beyond its instructions and model, an agent carries a few behavior settings, also in the Behavior section:
- Thinking — the model’s reasoning effort, from Off through Low, Medium, High, to XHigh. It’s an agent-wide setting: it decides how hard the agent thinks before answering, on every message, rather than something you toggle per request. Off disables extended thinking entirely. Higher levels trade speed and cost for more deliberate reasoning.
- Generation settings — the usual LLM call knobs (temperature, max output tokens, and so on), for fine-tuning how the model responds.
An agent can also be scoped to a set of tools — what it’s allowed to do beyond plain chat. That policy, and the skills an agent can draw on, are covered on their own page.
Custom agents {#custom-agents}
The New Agent chooser also offers Create custom agent — for an assistant you want to build from the ground up rather than starting from Le Chaton Fat. The form is the same one you’d use to edit any workspace agent, grouped into sections:
| Section | What you set |
|---|---|
| Identity | Name, an avatar, a short description, and tags. (The id is fixed once created and must match ^[a-z0-9-]{3,32}$.) |
| Provider & Model | The provider and model this agent runs on (both optional). |
| Behavior | The system prompt, the Thinking level, and the LLM config (temperature, max output tokens, top-p, stop sequences). |
| Tools | What the agent may do beyond chat — see Skills. |
Save, and the agent joins your workspace’s Agents tab alongside any you’ve cloned.
Your workspace’s default agent
“Default” means two related things, and it helps to keep them apart:
- Le Chaton Fat is the built-in you start from — the seed agent you clone into a workspace to get your own editable copy.
- A workspace’s default agent is the one it reaches for first — the agent new sessions begin with unless you pick another. Each workspace agent card carries three actions: a star to set it as the default, a pencil to edit, and a power icon to disable it.
The Agents tab always shows the current pick on a Default agent line. Until you set one of your own, it reads Le Chaton Fat (system agent) — a workspace with no default of its own falls back to the shipped default. The intent is that each workspace owns its default: you clone Le Chaton Fat (or build a custom agent), tune it for that workspace’s work, and star that copy — rather than every workspace leaning on one global assistant.
Disabling an agent
A workspace agent can be disabled — set aside without being removed. A disabled agent drops out of the pickers and wears a Disabled badge, but its files stay on disk, so you can re-enable it later. This is the way to retire an agent you no longer want cluttering the list; the record is preserved so nothing that referenced it breaks silently.
Where to find your agents
Everything to do with agents lives on the workspace’s Agents tab. The Available agents catalog at the top holds the read-only Shipped agents to clone from and the Custom tile to build from scratch; your own workspace agents sit below, where you clone, create, edit, star (set default), and disable them.
What to read next
An agent is the assistant; now for what it can reach and where it runs: