Agent Studio and Platform Services Diagram

Two Authoring Doorways

• Browser editors autosave through collab-gateway, which flushes config back to ai-manager and keeps durable Yjs state in collab_documents
• The Yjs converters are a fixed allow-list, so an unlisted top-level config key is silently dropped on every flush
• External clients use versioned REST and stateless MCP with fine-grained management scopes, never the platform-admin bypass

AI Authoring Is Sandboxed

• Generate with AI and the editor Assistant both run a scoped CLI inside workspace-agent-runner
• The sandbox writes only through the generation MCP endpoint using a per-job token bound to one target
• Validation is two tiers: structural JSON Schema in Studio, buildability preflight in ai-orchestrator

Publish Is A Database Fact

• Runtimes resolve the published agent_versions row themselves; Studio pushes no config to them
• Tenant isolation is Postgres RLS, while Authentik groups stay authoritative for membership and platform admin
• Not drawn: MinIO catalog bundles, CRM embeds, OTLP telemetry

Telephony Edge

• Inbound CCaaS is default-deny: ari-controller rejects calls without an enabled ccaas_pipeline_integrations row
• SIP calls land on Kamailio/Asterisk; ari-controller allocates media-proxy legs that stream gRPC audio to ai-orchestrator
• Janus, rtpengine, and coturn (WebRTC/TURN media plane) are folded into the SIP edge here