AGENTIC AI IN PRODUCTION — PHASE 1 · LAYERS 1–2 · THE FRONT DOOR
Agentic AI in Production — Phase 1 Channels, Identity & Gateways
One request from the moment it arrives on any channel to the moment it is allowed to become a plan: edge protection, one request envelope, authentication, authorization, guardrails, a metered model gateway, and a human decision when the risk demands one.
The RAG series' forces still apply. Agentic systems add a seventh: every action must answer who caused it, under whose authority, and whether that can be proven later.
1
Experience & Channels
many doors → one edge → one envelope → sync or async
Identity must travel with the channel. The adapter's job is to hand Layer 2 a verifiable envelope — not a shared service credential that quietly discards who the user was.
The gateway is critical-path infrastructure: multi-zone, config-as-code, fail closed on authorization and key custody, degrade-and-alert on optional analytics. No client may reach a model or tool without passing through it.
X
Cross-cutting Concerns
one correlation context feeds every concern, at every hop
Cross-cutting does not mean "someone else's job". Each concern is only as good as the context created at the edge: without a trace id, tenant and identity on every hop, none of the eight can be answered after the fact.
Phase 1 invariant: every request is authenticated on a known channel, authorized as the intersection of user and agent permissions, metered per team and token, screened before and after the model, attributable to a human or an explicit autonomous-agent policy, and — when risk demands — held for a human decision that is bound to the exact action shown.
Layer 1 · Channels
API Gateway / Identity
AI Gateway
Guardrails / deny paths
Approvals
Cross-cutting
⋯ dashed = supporting / conditional path
◆ diamond = decision point
Hover any box for its role and trade-offs. Moving dots trace the direction of execution through each band.