AI agents

Ship Builder agents inside your package — installed into the tenant's Agent Builder, governed by their keys, budgets and kill switches.

Declaring an agent script

"agentScripts": [{
  "name": "Property valuation assistant",
  "description": "Estimates listing value and drafts the summary",
  "systemPrompt": "You are a real-estate valuation expert. Given a property record …",
  "skills": ["create_task", "search_leads"]
}]

field	notes
`systemPrompt`	≤ 20k chars; this IS your agent — write it like production code
`skills`	names from the platform tool registry; the agent can ONLY call declared skills (scope violations are detected and flagged)

Governance your agent runs under

Tenant model keys — agents run on the tenant's own AI-provider key (Settings → AI Providers). No key, no run: the platform never silently bills its own.
Token budgets — every agent has a monthly token budget; exceeding it returns 429 BUDGET_EXCEEDED.
Kill switch — tenants and operators can halt agent tool-calls platform-wide, per class, or per agent in under 5 seconds.
Approvals — destructive tool calls can require human approval (two-person for critical classes).

Invoking agents from workflows

Pair an agent script with a run_agent workflow action to build closed loops: event → agent reasoning → CRM mutation. The action passes the event payload as context and DLQ-retries when the agent doesn't answer.

Write prompts that ground on the provided record context and refuse to fabricate. Review reads your systemPrompt — hallucination-prone prompts are the most common agent-script rejection.