Asset Management Agent Sandboxing How It Works | Moring
Agent sandboxing works through layered architecture: an isolated runtime forms the boundary, a mediated channel carries every request across it, per-agent identities and injected credentials establish who is acting, a policy engine evaluates each action before execution, and retained logs make the whole flow reconstructable. This page walks through the asset management-specific anatomy of that architecture.
Short answer
Agent sandboxing works through layered architecture: an isolated runtime forms the boundary, a mediated channel carries every request across it, per-agent identities and injected credentials establish who is acting, a policy engine evaluates each action before execution, and retained logs make the whole flow reconstructable. This page walks through the asset management-specific anatomy of that architecture.
How it works
Modern sandbox architectures share an identifiable anatomy, and understanding it helps asset management platform teams evaluate vendors and design their own deployments. Reference implementations from major infrastructure providers describe the same shape: a sandbox boundary containing the agent's runtime, a mediated channel through which every request travels, an authentication layer that establishes the agent's identity, and policy enforcement at the crossing points.
The boundary itself is built by the runtime. Depending on the tier, that runtime is a hardened container, a microVM, or an equivalent isolation environment, and its job is to make sure that code running inside it cannot directly reach the host, the network, or other systems. Isolation guidance from security teams emphasizes exposing narrowly defined interfaces rather than open-ended access, which is why a well-built sandbox feels restrictive to the agent by design.
The mediated channel is the load-bearing element. Every file read, network request, and tool call the agent makes travels through a mediation layer that authenticates the request, applies policy, and logs the outcome. Deterministic interception at this layer evaluates each action before execution, and in high-security deployments the mediation composes with the runtime isolation for defense in depth. For asset management, this channel is where restriction lists, data-residency rules, and approved-destination policies become enforced decisions rather than written intentions.
Identity and credentials flow through the same channel. The agent obtains its first credential from the boundary's authentication service, receives scoped, short-lived tokens for each destination, and never holds standing secrets. Renewal and revocation are handled by the boundary, which means that revoking an agent's access is an administrative action at the control plane, not a hunt through application code. Per-agent identity is what makes attribution across the estate possible.
The split between control plane and execution plane completes the design. The harness that owns the agent loop, model calls, approvals and recovery state runs in trusted infrastructure, while model-directed work happens inside the sandbox with its own filesystem, commands, ports and snapshots. Keeping those boundaries separate means authentication, billing, audit logs and human review never live in the container the agent can influence, and stateful work can pause for review and resume in the same workspace without re-provisioning anything.
Observability comes from the logs the channel produces. Because every crossing is recorded, the firm can reconstruct any workflow, measure cost per case, detect anomalous patterns, and produce the evidence packs that model risk management reviews under SR 11-7 and the SR 26-2 direction require. Termination controls close the architecture: kill switches and session pauses live in the control plane, outside the agent's reach.
The practical evaluation question for asset management firms is whether a proposed sandbox enforces policy at the mediated channel with per-agent identity and complete logging, or whether it merely promises isolation while letting the agent call systems directly. The former is governable; the latter is a boundary in name only.
Frequently asked questions
What are the main architectural layers of an agent sandbox?
Reference implementations converge on four layers: an isolated runtime that forms the boundary, a mediated channel that carries every file read, network request and tool call, an authentication layer that issues per-agent identities with short-lived scoped credentials, and a policy engine that evaluates each action before it executes. Retained crossing logs and control-plane termination controls complete the design, which is what gives asset management firms both reconstruction and interruptibility when an agent misbehaves.
Why does the mediated channel carry so much of the enforcement load?
Because that is where enforcement actually becomes real. Interception at the channel evaluates every action against policy before execution and records the outcome, so restriction lists, residency rules and approved destinations turn from documents into enforced decisions. Security guidance recommends composing this application-layer mediation with runtime isolation for defense in depth, which is the combination asset management workloads with fiduciary consequences need rather than either control alone.
How do sandbox credentials and revocation behave in practice?
The agent receives its first credential from the boundary's authentication service and then scoped, short-lived tokens per destination, never holding standing secrets inside the sandbox. Renewal and revocation are control-plane actions at the boundary, so cutting access during an incident is administrative rather than a code change or a credential hunt. Asset management teams get fast containment and clean attribution across the agent estate as a byproduct of the design.
How should asset management teams test a vendor's sandbox claims?
Ask whether policy is enforced at a mediated channel with per-agent identity and complete crossing logs, or whether the agent retains direct paths to systems despite the isolation promise, then request the log output from a deliberate violation attempt. The first design produces the evidence model risk reviews under SR 11-7 and the SR 26-2 direction expect; the second is a boundary in name only and will fail both attribution and examination tests when it matters.