# Build: agentic flows with the wiring showing
For AI-fluent builders. Autonomy model: L1 read-only (where we ship), L2 suggest, L3 act with approval, L4 autonomous.

Guardrails we start from: read-only credentials scoped per source; untrusted content never triggers actions; anything external needs a human; label uncertainty honestly; log every run.

Patterns: chief-of-staff agent, read-only first, 4-hourly watchdog, voice agent with handoff, email + portal ingest, knowing when to stop. Four are finished Blueprints (payment reconciliation, receipts to bookkeeping, delivery exceptions, content repurposing); the rest are in progress.