Reference architecture

Architecture of the governed action loop

A production agentic platform is more than a model with tools. Intent becomes a decision, a decision becomes an identity-scoped action, an action changes platform state, and telemetry feeds the next decision — with governance, observability, and human authority wrapping every step.

The governed action loop

Intent

A human or event expresses a desired outcome, not a procedure

Reasoning

The agent plans how to reach the outcome

Context

current system state

Memory

what worked before

Policy

what is allowed

Decision

A proposed action, justified and inspectable

Identity

Who is acting, on whose behalf, with what scope

Action

Execution through a platform capability, never raw infra

Platform

Golden paths, pipelines, and APIs reconcile state

Telemetry

Outcomes and action traces feed back into context

Governance

wraps every step

Observability

sees every step

Human authority

bounds every step

How to read this diagram

The spine is the governed action loop. An intent — a human request or a system event — enters the platform. The reasoning plane plans a route to that outcome, drawing on three inputs: context (current system state), memory (what worked before), and policy (what is allowed). Planning produces a decision: a proposed action that can be inspected before it runs.

Before anything executes, the identity plane answers who is acting, on whose behalf, and with what scope. Only then does the action plane execute — through platform capabilities (golden paths, pipelines, provisioning APIs), never raw infrastructure credentials. The platform reconciles state, and telemetry flows back into context, closing the loop.

Around everything run three cross-cutting planes: governance (the rules and approval workflows), observability (the traces of what was decided and done), and human authority (the escalation and approval model that keeps consequential actions in human hands).