Agents
One agent is one model with a soul, a set of tools, and a budget.
The soul
The system prompt, written as prose. It is the whole of what the agent is told about itself.
Model and provider
The catalogue has a card for every model, read from models.dev: reasoning, tool calls, vision, context. A model new to models.dev arrives switched off, and the platform's operator switches it on; only switched-on models can be picked.
Budgets
Temperature: the provider's default (the agent sends none), or a custom value, offered only for models that take one. Max tokens: 4096 by default, 256 to 200000. Tool rounds: 10 by default, 1 to 100. Context pruning: off, light or aggressive.
What an agent can reach
The skills, plugins and MCP servers wired to it in Garage → Agents → Tools (see Skills, plugins and servers), the memories and libraries wired to it, and other trons named under Cross-tron: the agent can hand them a sub-task.
Running on a schedule
A cron set on an agent in the Lab is a row of the one schedule table (see Schedules in the API part). There is no second clock inside the agent.
Libraries and context files
An agent can be given libraries of documents, to read or to read and write: it searches them with search_documents and reads one with read_document, and every answer it takes from them carries a citation. Garage → Agents → Context files is the short way: drop files on the agent, and they go into its library (made with the defaults if it has none) and are indexed. See Libraries and documents.