Skip to main content

AgentSpec

Type Alias AgentSpec 

Source
pub type AgentSpec = Spec;
Expand description

App-facing alias for the domain agent specification used by workflow automation.

A spec is the stable rules an agent runner receives before it builds a prompt packet: the workflow identity, business purpose, read/draft tools it may use, actions it must never take directly, and deterministic review gates that keep resort staff in control of bookings, messages, schedules, and safety-sensitive decisions.

Aliased Type§

pub struct AgentSpec {
    pub name: Name,
    pub purpose: Purpose,
    pub allowed_tools: Vec<ToolName>,
    pub forbidden_actions: Vec<ForbiddenAction>,
    pub default_review_gates: Vec<ReviewGate>,
}

Fields§

§name: Name

Stable workflow-facing agent name.

This is an identifier such as manager-daily-brief or booking-triage, not a human display label. It connects prompt packets, outputs, and audit evidence back to the spec that constrained the agent run.

§purpose: Purpose

Business purpose for the agent’s draft or evidence work.

The purpose should state the resort operation being supported, such as summarizing labor risk, drafting customer-safe follow-up, or routing vaccine-document ambiguity; it does not grant live-action authority.

§allowed_tools: Vec<ToolName>

Tool names the runtime may expose to this agent.

These should be read-only, draft-only, or task-creation surfaces scoped to the workflow. They are the positive capability list for context building, not permission to bypass review gates or mutate source systems.

§forbidden_actions: Vec<ForbiddenAction>

Live or unsafe actions the agent must not perform directly.

Examples include confirming bookings, promising availability, changing labor schedules, waiving deposits, diagnosing pets, or sending customer messages without approval.

§default_review_gates: Vec<ReviewGate>

Human or deterministic app review gates required for the workflow.

These gates keep manager approval, customer-message approval, medical document review, and similar authority outside the model-generated output.