The brain and model tiers

The brain

The brain is the part of Crewdesk that decides what to do. It is a conversational CrewAI Flow with one task agent, deployed on CrewAI AMP. It is the same brain for every user and every assistant; what changes from one turn to the next is the context it gets.

The brain never holds credentials. When it needs data or an action in one of your apps, it asks the Hub (a Hub call), which checks permissions and policy and runs it. When it needs your computer, it asks your client (a client call), which runs the tool on your Mac. See Turn lifecycle.

The Hub can also run the brain itself. If the brain on AMP fails before it has shown you anything, the Hub runs that turn locally, so a hiccup on one side doesn't cost you the answer.

Model tiers

Each turn runs on one of three tiers:

Tier	Used for
light	A small, fast model. Only when you pick it.
default	The everyday model.
strong	A large model, for design, frontend work and hard requests. A strong turn also drafts its pages and builds with it.

With auto (a new chat's setting), the brain picks the tier per message: the strong tier for design, frontend and hard code, the default tier otherwise. Your choice for the chat comes first: the app's model selector, /model in a chat, or --model in print mode. An assistant can set its own default tier.

If a model's key is refused, the tier steps down by itself and the reply says so.

Specialists and helpers

Some jobs go to agents of their own, which the task agent calls as tools:

• Data analyst: answers a data question through one of your query automations, reading the schema and samples first.
• Automation diagnosis: investigates why one of your automations failed, from its runs and traces.
• Web research, file analysis and a CrewAI expert: helper agents on AMP for deeper research, questions about attached files and spreadsheets, and CrewAI questions beyond the docs index.