Agent tools & permissions
An agent’s instructions describe desired behavior. Runtime permissions determine what the agent can actually access or do.
Depending on the workflow, an agent can be eligible for tools such as:
- web search
- URL reading
- Project retrieval
- Knowledge retrieval
- calculation
- repository search
- FORGE filesystem/build tools
- validators
- artifact operations
Tool availability depends on plan, model capability, workflow, organization policy, and administrator configuration.
Least privilege
Give the agent only the tools needed for its role.
A documentation agent may need Project/Knowledge retrieval and search. It usually does not need destructive build filesystem operations.
A FORGE-oriented agent may need repository and validator access but does not automatically need billing administration.
Agent instructions cannot grant permissions
Writing “you have access to all organization files” in the prompt does not create access. The runtime still checks the active user, organization, project, resource, and tool policies.
Project and Knowledge scope
Attach the narrowest relevant context. An agent used for one product should not automatically retrieve from every organization knowledge base.
Sensitive tools can require approval before execution. An agent can propose the action, but the approval system decides whether execution may continue.
Model compatibility
A selected model must support the required tool behavior. A fallback route that cannot use the tool is not equivalent for that task.
Review after role changes
When an agent’s job changes, review its tools and context. Old permissions often survive longer than the instructions that originally justified them.
Never place API keys, passwords, or private credentials into agent instructions as a substitute for proper tool/integration configuration.
