Checking review mode… Reading PRODUCTION_STRICT from governance API health.
Glass-box Arch · pipeline · product Operate Onboard → Gateway → Monitor Architecture Stack · tradeoffs · live ops
Governance · glass-box
Watch policy before side effects Left: stack + live metrics. Center: replay enforcement_steps from POST /api/gateway/tool-request (post-response, not SSE). Right: run the governed intercept.
Architecture
How AegisAI is wired Governance control plane for agent side effects: policy, HITL approvals, execution tokens, and signed audit — kept separate from orchestration so agents cannot bypass the gate.
Every dangerous tool call in the vpeetla-ai stack is meant to route through AI Gateway before production impact.
Tool plane · AI Gateway Core product — intercept deploy / notify / refund, issue tokens, HITL, kill switch, audit.
Model plane · LLM metrics Optional satellite — read-only cost & cache metrics from aegis-llm-gateway + semantic cache (not tool intercept).
Stack
Control plane layers L1 Control Room Operator UX
Operate journey HITL queue Incidents / kill switch LLM metrics L2 AI Gateway Side-effect gate
Tool request API Policy / OPA Approval tasks Execution tokens L3 Execution broker Idempotent connectors
HTTP connectors Deploy / notify adapters Rollback refs Audit events L4 Ops & state Regulated persistence
Control-plane DB /api/v1/ops/metrics Observability adapters Principal tradeoffs
Decisions with explicit costs Governance separate from VAP orchestration Gain One policy engine for all agents across products
Trade Cross-service latency on gated actions
SQLite / Postgres control plane as source of truth Gain Audit + HITL state survives trace TTL
Trade Not a serverless-only design
Tool AI Gateway ≠ LLM metrics plane Gain Clear product story; model routing stays in satellite services
Trade Two surfaces operators must not confuse
MCP governance proxy Gain Tool calls gated like HTTP side effects
Trade Proxy hop on every tool invocation
Record
ADRs, case studies, SLOs Live ops
Production metrics Loading live ops metrics…
AI Gateway · enforcement replay awaiting run
Run a tool intercept to animate authorize → risk → policy → allow / HITL / block.
Governed tool call 1. Agent requests deploy.vercel_release → 2. Gateway evaluates → 3. Allow, HITL, or block with audit.
Tenant bank-demo
Agent agent-fe-builder
Tool deploy.vercel_release
Target vercel Run governed intercept