What an agent actually is

An agent is a saved configuration: instructions, a model, a set of tools, and a few limits. It is not a separate program and it does not hold data of its own.

Where its abilities come from

When you connect software, AgentOS reads what that software can do — endpoints, tables, MCP tools — and turns each capability into a tool with a name, a description and a schema for its arguments. An agent's abilities are exactly the tools you gave it, and nothing else.

That is why an agent cannot "look something up on the internet" or "run a quick query" unless a connection you made exposes a tool that does it.

What the model is and is not

The model decides *what to propose*. It never decides what happens. Every call it suggests is checked independently:

• •the tool exists, and this agent has it
• •your organization owns the connection behind it
• •the arguments match the tool's schema
• •the person the agent is acting for has permission
• •policy allows it, or a human is asked first

If any check fails, the call does not run, and the model is told why so it can correct itself or explain.