Where Does Geordie Sit?

Most teams already run EDR, identity and cloud controls, and some have added a model gateway or secure MCP. Here’s where Geordie plugs in. We believe it is the only tool in the market that understands agents as systems, not surfaces, and gets you the understanding you need to take control of operations.

Showing stack for

1The agent calls a model, through a gateway where you run one.
2The agent calls tools over MCP, through secure MCP where you host servers.
3Direct routes. Third-party harnesses, shadow AI and CLI tools often skip the gateway and MCP.
4Geordie instruments each harness and applies controls inline, so telemetry and controls travel with the agent.
5EDR sensors report to the platform and the SOC. The audit trail feeds the same SIEM.
6Coding agents push to repositories.

Trust boundary
Containment layer
Stack layer
Context layer, position to follow
Geordie works across this
You choose: open source, hyperscaler or vendor
Direct route
Telemetry and controls

Where we land

Four positions shape the diagram above. Each one follows from how agents operate.

Add scaffolding where you want it

Sandboxes and host controls are optional. If you want a sandbox, use what your hyperscalers provide on AWS and Microsoft, or what your developers already run in Docker. If you want higher fidelity or zero trust on the host, we recommend nono or Minimal. Geordie works across Mac, Windows and Linux with or without them.

Keep EDR and XDR as the detection backbone

CrowdStrike leads on threat detection and Defender leads on extensibility and cost. Both already carry heavy signal for the SOC, so agentic specialism belongs with a partner who works in that vocabulary.

Treat gateways and secure MCP as optional

A gateway earns its place for model routing preferences, and we recommend LiteLLM. Secure MCP earns its place when you host and manage a wide swathe of your own MCP infrastructure, and we recommend Bifrost or the Microsoft MCP Gateway. Shadow AI, other harness types and tools where MCP is not central sit outside both.

Give agent behavior a layer of its own

Agents act continuously, so observability, monitoring and controls have to run at their pace. Agentic Security & Operations runs alongside the agent and covers every harness in the estate. Geordie is our solution of choice.

Three operating models

Cloud agent factory and ops

Every agent in this model is first-party and runs in the cloud. The platform controls the runtime, the identity and the network, so containment comes from turning on what the hyperscaler already provides. The open question is behavior across a fleet that changes daily.

Who this is for: Platform engineering, SRE and security teams that build and run agents at scale in their own cloud accounts.

What it looks like

The factory builds first-party agents on vendor ADKs and SDKs, open source frameworks and custom harnesses, and ships them to managed runtimes.
Hyperscaler isolation or Docker containers contain each session.
A gateway is useful for model routing. Secure MCP is useful if the factory hosts many internal MCP servers.
The fleet layer observes and controls every deployment as it runs.

Core Plan for these first.

Agentic Security & Operations Fleets of cloud agents change daily. This layer gives platform and security teams the same view of every deployment, with controls applied as agents run. Cloud agent platforms The primary surface. Everything in the factory runs here. First-party agents and custom harnesses Custom harnesses and first-party agents on ADKs, SDKs and open source frameworks are what the factory produces and runs.

Recommended Add these as the estate grows.

Model gateway Routing and spend control matter at factory scale, where you own the agents' base URLs. EDR and XDR Workload protection for the hosts and clusters under the factory. Native controls Cloud platform policy, identity and guardrails are the native baseline for the factory.

Optional Adopt these when a specific need appears.

IDEs and code repositories Teams that build the factory use IDE agents, and repositories hold the agent definitions. Browser agents Browser automation appears in some factory workloads. Secure MCP Adopt it when the factory exposes a large internal tool catalog through MCP. Sandbox Worth adding for untrusted code and multi-tenant factories, where each session benefits from its own boundary. Operating system Minimal environments and small images for the hosts and containers under cloud runtimes, where you want a second boundary.

Rarely used These layers play a small part in this model.

Desktop harnesses A cloud-only factory does not rely on desktop apps.

What to deploy, layer by layer

Agentic Security & Operations

Cloud agent platforms

Desktop harnesses

IDEs and code repositories

Browser agents

First-party agents and custom harnesses

Model gateway

Secure MCP

Sandbox

Operating system

EDR and XDR

Native controls

Fleet Geordie

Agentic Security & Operations

Agents move continuously, so Agentic Security & Operations runs alongside them.

Geordie owns this layer Geordie owns this layer. It discovers, maps, observes, assesses and controls every agent, and it sits in the path of each agent so controls move at the same pace.

Cloud agent factory and ops Core Fleets of cloud agents change daily. This layer gives platform and security teams the same view of every deployment, with controls applied as agents run.
Corporate agentic usage Core Employees adopt harnesses whether security has approved them or not. Discovery and behavior data start here.
Multi-national organization Core One operating picture across first-party agents, third-party harnesses, code and endpoints, with a single audit trail.

What is this?

Agentic Security & Operations gives security and governance teams one operating picture of every agent in the organization. It discovers each agent, maps the tools and permissions it holds, observes how it behaves over time, assesses the risk that creates and applies controls.

Geordie is the System of Authority for Agentic Security & Operations. It deploys directly to the agent across cloud, code, and endpoint, and it is vendor-agnostic.

What is its purpose in the stack?

Every other layer in this stack sees a slice: a host, a process, a request, a tool call. An agent is a system that spans all of them.

Risk in an agentic estate builds up across sequences of actions and the context that flows between steps. Seeing it takes a layer that sits with the agent and watches behavior as it happens.

This layer works with the scaffolding around it: sandboxes bound what an agent can reach, and the fleet layer shows what the agent does inside those bounds and shapes it inline, through the agent's own context and configuration.

What should I use?

Geordie, which covers the harnesses your developers and corporate users already run, and the agents you build yourself, from one place.

Beam, Geordie's remediation engine, applies deterministic controls inline. The immutable audit trail feeds downstream into your SIEM.

Open standards help you prepare the ground. OpenTelemetry's GenAI conventions give your own agents a shared vocabulary for traces, and open source tools such as Langfuse and Arize Phoenix record what a single application did. A fleet layer extends that view across every harness in the estate.

References you can deploy

Surfaces Context layer · position to follow

Cloud agent platforms

Geordie works across this layer Geordie deploys to the agents running on these platforms. The runtime is yours or your hyperscaler's, and Geordie works across all of them.

Cloud agent factory and ops Core The primary surface. Everything in the factory runs here.
Corporate agentic usage Rarely used Corporate usage is led by desktop harnesses, so cloud platforms appear only where teams build their own.
Multi-national organization Core First-party agents for internal use and for customer-facing products run here.

What is this?

Agents that your teams build and operate in cloud accounts. They run on managed agent runtimes from the hyperscalers, on Kubernetes-based platforms or on workflow engines. An agent factory is the pipeline and platform that produces and runs them at scale.

What is its purpose in the stack?

This is where first-party agents live and where the most direct controls exist, because you own the runtime, the identity and the network.

It is also where agent counts grow fastest, which makes the factory the right place to build in observability from the first deployment.

What should I use?

Start with the managed runtime your hyperscaler provides, such as Amazon Bedrock AgentCore, Azure AI Foundry or Google's Gemini Enterprise Agent Platform. Use Kubernetes with a workflow engine where you need portability. Choose the framework in the First-party agents and custom harnesses layer.

References you can deploy

Traffic-boundary controls LiteLLM, when routing matters

Model gateway

Useful when model routing preferences matter to you, though not essential.

Your choice Your choice. We recommend LiteLLM, and you can run any open source or commercial gateway. Geordie covers agent behavior beyond the traffic a gateway sees.

Cloud agent factory and ops Recommended Routing and spend control matter at factory scale, where you own the agents' base URLs.
Corporate agentic usage Optional Desktop harnesses rarely let you set routing, so gateway coverage is partial.
Multi-national organization Recommended Useful for first-party agents and products. Plan for parts of the estate that route around it.

What is this?

A model gateway is a proxy between agents and model providers. It offers one endpoint, routes calls across models and providers, holds provider keys, enforces budgets and records requests.

What is its purpose in the stack?

It gives platform teams control over model routing: which model serves which workload, what each team may spend and where traffic fails over. It works where you control the client's base URL, which is typical for agents you build.

Third-party harnesses, shadow AI and tools where the model is one component often reach providers directly or through their own routes. A gateway sees the traffic pointed at it, so it describes part of the estate and does not show what an agent does with the answers.

What should I use?

Adopt a gateway when model routing preferences matter to you: cost, latency, provider choice or failover. We recommend LiteLLM, an open source proxy you run yourself, and we look to the fleet layer for behavior.

References you can deploy

Containment Optional · nono or Minimal

Operating system

An additional, optional control mechanism for higher fidelity or zero trust. Where you want it, we recommend nono or Minimal.

Your choice Your choice. We recommend nono or Minimal, both open source or third party. Geordie works with or without them.

Cloud agent factory and ops Optional Minimal environments and small images for the hosts and containers under cloud runtimes, where you want a second boundary.
Corporate agentic usage Optional nono on corporate laptops gives each harness process kernel-level limits, for teams that want zero trust on the endpoint.
Multi-national organization Optional nono on endpoints and developer machines and Minimal or small images in the cloud, applied where the risk justifies them.

What is this?

The operating system layer is the kernel, filesystem permissions, network access and credentials on the machine or image an agent runs on. Everything an agent does ends in a system call.

Host-level controls are an additional, optional mechanism. Geordie works across Mac, Windows and Linux without them. Add them when you want higher fidelity of control or a zero trust posture on the host.

What is its purpose in the stack?

Limits applied at the kernel hold below the agent and below any guardrail inside its process. A small image reduces what an agent can reach on a host in the first place.

They add depth to the other layers. Observability, monitoring and controls at the agent layer continue to do the main work, and host controls give a second, independent boundary beneath them.

What should I use?

Where you want host-level controls, we recommend nono or Minimal.

nono confines agent processes with kernel mechanisms: Landlock on Linux and Seatbelt on macOS, with Windows supported through WSL2. It scopes authority per tool call and injects credentials through a proxy outside the sandbox.

Minimal builds isolated, reproducible environments from source. Each runs in a microVM with file system and network policy, and each package ships with an SBOM. It runs locally on Linux and macOS and is in public beta, so check its current status before you plan around it. For hosts and containers in your cloud, distroless images give a similar small footprint today.

References you can deploy

Detection Keep, and add a specialist

EDR and XDR

Keep EDR and XDR as the detection backbone, and pair them with an agentic specialist.

Your choice Your choice. Keep CrowdStrike or Defender. Geordie is the specialist partner for agentic operations that sits beside them.

Cloud agent factory and ops Recommended Workload protection for the hosts and clusters under the factory.
Corporate agentic usage Core The detection backbone for every laptop that runs a harness.
Multi-national organization Core The shared detection backbone across regions and business units.

What is this?

Endpoint and extended detection and response platforms watch hosts, identities and cloud workloads for threat behavior and respond when they find it.

What is its purpose in the stack?

They are the detection backbone of the security operation, and analysts already work in them. In an agentic estate they still catch malware, credential theft and lateral movement on the hosts where agents run.

These platforms carry a great deal of signal from many sources, and agent behavior is one more stream. It has its own vocabulary of tool sequences, context flows and delegated authority. Agentic specialism belongs with a partner who works in that vocabulary and passes evidence downstream to the tools your analysts already use.

What should I use?

CrowdStrike is the usual pick when threat detection depth leads and Microsoft Defender when extensibility and cost lead, so choose on those two axes and then add the specialist layer for agents.

References you can deploy

Native controls Use, then verify

Native controls

Native controls set the baseline. Oversight, assurance and additional controls remain necessary on top of them.

Geordie owns this layer Geordie provides the oversight and assurance across the native controls of every harness, and adds controls where native ones run out.

Cloud agent factory and ops Recommended Cloud platform policy, identity and guardrails are the native baseline for the factory.
Corporate agentic usage Core Desktop harnesses depend on vendor enterprise plans and managed configurations for their controls.
Multi-national organization Core Many products at once means many sets of native controls to deploy, check and keep aligned.

What is this?

Native controls are the settings and policies that agent products, cloud platforms and operating systems ship with. They include permission rules, managed configurations, admin consoles, data policies and audit logs.

They usually sit in enterprise tiers, and many take effect only when an administrator deploys them as a managed configuration. A default install carries less than a configured one.

What is its purpose in the stack?

They are the first line of control. The vendor enforces them, they sit closest to the product's own behavior, and they cost little to adopt once you hold the right plan.

Each set covers one product, in that product's vocabulary, on that vendor's roadmap. An estate with several harnesses has several sets of native controls, with gaps between them and settings that drift over time.

Three needs continue alongside them: oversight, which gives one view of what agents do across every product; assurance, which gives evidence that the configured controls are in force and working; and additional controls that act on agent behavior inline, independent of any one vendor.

What should I use?

Buy the enterprise tier of the agent products you standardize on, and deploy their managed configurations centrally through MDM or cloud policy. Treat the vendor settings as your baseline.

Then verify them. Geordie provides oversight and assurance across the native controls of every harness in the estate, and adds controls that operate at the pace of the agent where native controls run out.

References you can deploy

Take control of AI agent uncertainty

Get a single source of truth for every autonomous agent your organization runs.

Book a demo

Where Does Geordie Sit?

Where we land

Add scaffolding where you want it

Keep EDR and XDR as the detection backbone

Treat gateways and secure MCP as optional

Give agent behavior a layer of its own

Three operating models

Cloud agent factory and ops

What it looks like

Corporate agentic usage

What it looks like

Multi-national organization

What it looks like

What to deploy, layer by layer