CONCEPT · AGENT
Agent
An autonomous AI worker with a role, an authority envelope, and an auditable trail.
What it is
An agent is a named reasoning process bound to a role ("triage", "scanner", "scribe"). It has an explicit authority envelope declaring which connectors and actions it can invoke, and every call it makes is recorded as a Decision.
Why it exists
Generic "AI assistants" are unmanageable at scale. An agent has scope, budget, and provenance — you can answer "who decided this?" for every automated action in your business.
How it runs
Agents are executed inside workflow nodes (ai-decision) and inside incident triage. The Lovable AI Gateway routes them to Gemini 3 Flash with strict schemas. Streaming markdown is emitted to the console in real time.
Shape
| Field | Type | Note |
|---|---|---|
| id | string | Stable identifier (obsidian, sentinel, scribe...). |
| role | enum | triage | scanner | scribe | compiler |
| envelope | string[] | Allowed connector.action pairs. |
| model | string | Model routed via Lovable AI Gateway. |