Infrastructure

The Pranix Agent Engine is a sovereign control plane for multi-product AI orchestration. Every operation is auditable, every mutation is protocol-governed, every decision is logged.

Control Plane

Central Supabase project managing task orchestration, execution memory, audit logging, and project registry across all products.

  • Task queue with DAG dependencies
  • Execution memory for cross-session continuity
  • MCP gateway with 29+ deployed tools

Worker Topology

Three-tier worker architecture: lightweight cron ticks, Supabase edge function workers, and Fly.io browser automation.

  • Tier 0 — Vercel cron (60s tick)
  • Tier 1 — Supabase heavy-worker (2min tick)
  • Tier 2 — Fly.io browser worker (planned)

Inference Cascade

Hybrid inference routing: deterministic rules first, then local models, then premium APIs. Cheapest successful result wins.

  • T0 — Deterministic (always available)
  • T1 — Ollama / NVIDIA NIM
  • T2 — Anthropic / OpenAI
  • T3 — Browser fallback (experimental)

Governance Protocols

Six versioned protocols governing all mutations: repo patches, hotfixes, CI failures, deployment verification, browser tests, and rollbacks.

  • Never push to main directly
  • All PRs require founder merge
  • Agents cannot revert autonomously

Event Sourcing

Every task state transition emits a task_event. Every MCP tool call is logged to mcp_audit_logs. Every deployment is verified.

  • Task event trail per operation
  • MCP audit log with latency + status
  • Deployment verification protocol

Security Model

Scoped bearer tokens, time-bounded grants, tool-level permissions, founder-only write access. No admin shortcuts.

  • Bearer to client_id to permissions
  • Grants require explicit approval
  • All writes audited to mcp_audit_logs