Asset Mgmt Agent Orchestration Internals | Moring
Under the surface, agent orchestration runs on a control-plane architecture: a policy and coordination plane that decides what agents may do, and a data plane where agents execute. Reference architectures divide responsibilities across planes for identity, policy projection, evaluation pipelines, and enforcement points. Asset managers evaluating orchestration should understand this internal structure, because it determines where their controls actually attach.
Short answer
Under the surface, agent orchestration runs on a control-plane architecture: a policy and coordination plane that decides what agents may do, and a data plane where agents execute. Reference architectures divide responsibilities across planes for identity, policy projection, evaluation pipelines, and enforcement points. Asset managers evaluating orchestration should understand this internal structure, because it determines where their controls actually attach.
How it works
Beneath the workflow diagrams, agent orchestration is a layered architecture with a governance-critical division of labor. Enterprise security was built to govern data boundaries, with controls governing crossings of a protected surface; production agents dissolve that assumption, because an agent reads context, calls tools, invokes connectors, and modifies systems of record across the enterprise. Reference architecture work responds with a multi-plane design: a control plane that decides what agents may do, and the runtime surfaces where governed requests are evaluated and executed.
The control plane holds policy, identity, and coordination state. Architectural descriptions of agent policy fabrics describe two governing mechanisms: signed policy projected into the runtime's own controls, so a sandbox is built from policy the enterprise approved, and per-request decisions, so a call the sandbox allows can still be denied or edited by the policy engine. For an asset manager, this division matters because it locates authority: the firm's policy, expressed once, governs every agent, tool, and integration attached to the fabric, rather than being re-implemented per agent.
Reference architectures for multi-agent systems describe the same structure in deployment terms: design patterns, implementation approaches, and governance controls for enterprise-grade systems, with open protocols such as MCP and A2A connecting agents and tools across vendors. The protocol layer matters practically, because an asset manager's estate will mix internally built agents with vendor agents, and orchestration that speaks standard protocols can govern both uniformly. Five-plane reference architectures go further, defining planes for runtime governance across identity, policy, observability, and enforcement, reflecting how production governance distributes across concerns.
Enforcement points are where the control plane meets execution. The architectural pattern attaches enforcement to the boundaries agents cross: tool invocations, data access, system-of-record writes, and inter-agent hand-offs. Each governed request is evaluated against policy at those points, with allow, deny, edit, or escalate decisions recorded as evidence. Handbook-level analysis of the AI control plane observes that control planes compete for authority over identity, orchestration, and data, and that many agentic projects are expected to be canceled when cost, value, or risk controls fall short, which makes the placement of enforcement an economic decision as well as a security one.
Identity runs through every plane. Agent identity frameworks describe a class of identity that does not fit traditional access management: agents authenticate programmatically, execute at machine speed, and operate across trust boundaries spanning systems and vendors. In the orchestration architecture, identity is what binds a request to a registered agent, a policy scope, and an evidence trail; without it, enforcement decisions have no subject and audit trails have no actor. Asset managers should therefore evaluate orchestration candidates on identity integration first, then on workflow features.
The practical reading for an asset manager is to map candidate or in-house orchestration against this structure: where policy lives, where it is projected, where enforcement points attach, how identity binds requests, and where evidence is produced. Governance toolkit guidance notes that enterprise AI is shifting from chat-based copilots to agents that reason, plan, use tools, and coordinate with other agents, creating a governance gap between promise and reality. Understanding the internals is how a firm closes that gap deliberately, choosing the architecture whose authority boundaries align with its fiduciary and supervisory obligations.
Frequently asked questions
What is the control plane in agent orchestration architecture?
The layer that decides what agents may do: it holds policy, identity, and coordination state, separate from the data plane where work executes. Agent policy fabric architectures describe two mechanisms: policy projected into runtime controls, so sandboxes are built from approved policy, and per-request decisions that can deny or edit a call even when a sandbox allowed it. Asset managers should map candidates against this structure, because the location of policy authority determines whether the firm's own governance or a vendor's defaults governs agent behavior across the estate.
Why do enforcement points matter in orchestration internals?
Because they are where authority becomes behavior. Enforcement points attach to the boundaries agents cross, tool invocations, data access, system-of-record writes, and inter-agent hand-offs, and each governed request is evaluated there with allow, deny, edit, or escalate decisions recorded as evidence. Since production agents cross data boundaries continuously, security based on perimeter controls alone does not hold. For an asset manager, the density and placement of enforcement points determine whether an out-of-mandate action is stopped at the boundary or discovered after client impact.
How does identity function inside the orchestration architecture?
Identity binds each request to a registered agent, a policy scope, and an evidence trail. Agent identity frameworks describe identities that authenticate programmatically, execute at machine speed, and span trust boundaries across systems and vendors, which traditional access management was not built for. Without per-agent identity, enforcement decisions have no subject and audit trails have no actor. Asset managers should evaluate orchestration candidates on identity integration before workflow features, since every other control, scoping, approval, evidence, and containment, attaches to the agent's identity.
Do open protocols change how orchestration internals work?
They change the connectivity layer, not the governance structure. Reference architectures describe multi-agent systems built with open protocols such as MCP and A2A connecting agents and tools across vendors, which lets an asset manager govern internally built and vendor agents uniformly through one orchestration fabric. The control plane, enforcement points, and evidence production remain architectural responsibilities of the layer regardless of protocol. Firms adopting protocol-standard estates should still verify that the orchestration layer evaluates every governed request itself rather than deferring to vendor-side policy.