Skip to main content

The harness matters more than the model

The agents your teams build themselves are often the most valuable in the estate and the least visible to off-the-shelf controls. Geordie understands them through the harness they run in, from the code that defines them to the work they do in production, and applies controls through the agent's own mechanisms without asking you to rebuild anything.

Frameworks and building blocks in this category

  • LangChain
  • Google ADK
  • AutoGen
  • MCP
  • Anthropic
  • OpenAI
  • OpenTelemetry
  • And more

Strategies tied to a single model will not age well

Large language models are not the only agentic future, and agent collaboration is already moving from fixed agent-to-agent patterns toward short-lived teams that form around a task and dissolve when it is done. Whatever model sits underneath, the harness, meaning the framework, SDK or platform that gives an agent its structure, is where its behavior is decided and where resilience needs to attach.

Approaches that sit outside the agent treat every harness the same way, because they only see the traffic that leaves it. Geordie reads the agent from the inside, through its configuration, context and telemetry, which is why it can tell a harmless tool call from a dangerous one in an agent your team wrote last week.

No rebuilds Geordie connects to custom agents without changing how your teams build them.

Four ways in, one record out

APIs and telemetry

Custom agents and harnesses in AWS, Azure and GCP connect through APIs and direct telemetry, including containerized environments.

Repository scanning

Agent configuration, tools, skills and dependencies are mapped in the repository, so risk is visible before an agent is deployed.

Optional lifecycle hooks

Where it is the better fit, a custom agent can also be added through its lifecycle hooks. It is one option, not a requirement.

Beam inside the harness

Controls are delivered through the agent's own mechanisms, such as lifecycle hooks, plugins and MCP, rather than a proxy in the path.

What each architecture can see of an agent you built

Geordie works alongside gateways and endpoint tools. The difference is where each one sits relative to the agent.

Gateway or proxyEndpoint-only sensor Geordie
Agents across cloud, code, endpoint and browser Partly Partly Yes
Skills, extensions, plugins and packages No Partly Yes
Risk found before an agent is deployed No No Yes
Behavior in the context of the whole agent No Partly Yes
Controls without a chokepoint in the execution path No Partly Yes

From the first commit to production

Before deployment

Find the risk in the code, not the incident

Repository scanning maps what a custom agent will be able to do before it runs: the tools it connects to, the skills and packages it pulls in, and the MCP servers it trusts. A vulnerable dependency or an over-broad tool connection is a finding in review rather than a surprise in production.

Owkin uncovered three critical CVEs its existing stack had missed, including an MCP command injection affecting 40 agents.

Relationship graph

One graph from the agent to the outcome

Geordie connects each custom agent to its owner, identity, model, tools, data and permissions, and to the work and actions that follow, tracking how access changes over time. Handoffs between agents are monitored, and relationships nobody designed are flagged.

Remediation, powered by Beam

Deterministic controls, delivered in the harness

Beam shapes behavior through context engineering delivered inside the agent, and policy is deterministic: no model judges whether an action is allowed, so a control behaves the same way every time. Responses escalate with the risk, from Monitor and Guide to Confirm and Enforce.

“A proxy approach introduces an additional third-party element which could add an additional layer of complexity. I've seen products in the market where if there's something wrong with the proxy, it can just open the metaphoric floodgates. Geordie's approach is very simple to connect. You can see exactly what's being connected. There's less of a dependency chain.”

Leo Cunningham CISO, Owkin

One record across every kind of agent

Most estates run all five. Geordie covers them in a single inventory.

Questions about custom agents and harnesses

You don't need to change how they're built, because Geordie connects through APIs, telemetry and repository scanning without any agent rebuilds, and lifecycle hooks are an optional way to add a custom agent rather than a requirement.
Geordie covers custom agents and harnesses running on your own infrastructure across AWS, Azure and GCP, including containerized environments and agentic software delivery loops. If you tell us which frameworks you build on, we'll confirm coverage for each one.
Beam reaches a custom agent through the agent's own mechanisms, such as lifecycle hooks, plugins and MCP, rather than through a proxy sitting between the agent and its tools.
Policy creation and enforcement are deterministic, so a control on your agent behaves the same way every time, and Geordie keeps AI for analysis and for understanding behavior.
Repository scanning maps an agent's configuration, tools, skills and dependencies in the code, so vulnerable components and over-broad access show up in review before the agent reaches production.
Geordie holds SOC 2 Type 2 and ISO/IEC 27001 certification and is GDPR compliant, and no third-party LLM sub-processors handle customer data.

See what your custom agents can reach

We will show you the agents your teams have built, what they connect to, and the risks already sitting in their repositories.