Important Guardian Agents and Agent Contracts are available in private preview. The screens, terminology, and behavior described in this section are based on pre-release designs and are subject to change before general availability.
As AI agents move into production, they can access sensitive data, call tools, and take actions autonomously, often faster than deployment-time policies or periodic reviews can keep up with. Guardian Agents extend control into runtime by continuously monitoring agent behavior against the rules set in an agent contract, and responding when an agent moves outside its established boundaries, from flagging and escalating an issue to redirecting or blocking the action.
- Agent Contract
- A governed, versioned asset that defines, in natural language, what an AI agent is allowed to do, what data it can access, and what outcomes it is responsible for. An agent contract turns informal expectations about agent behavior into a reviewable, auditable set of rules.
- Guardian Agent
- A runtime system that monitors an AI agent's actions against its Agent Contract, and intervenes, by denying, escalating, flagging, or rewriting an action, when that action violates the Agent Contract. Guardian lets you investigate and adjust the behavior of a single flagged agent without halting all other operations, containing risk while keeping the rest of your AI landscape running.
An agent contract can exist on its own, as a record of intended agent behavior. Guardian is what turns that record into active, real-time enforcement: without an agent contract, there is nothing for Guardian to enforce; without Guardian enabled, an agent contract is descriptive only and not enforced at runtime.
How Guardian connects to an agent
Guardian attaches to an AI agent in one of two ways:
- As an SDK installed directly in the agent's code, the primary path for agents you can modify directly.
- As a Gateway, an HTTP proxy compatible with major LLM provider APIs, for agents you can't modify directly, such as legacy systems, third-party agents, or workflow tools.
Either way, Guardian works from an Agent Fingerprint, a technical profile of the agent's tools, models, and prompts that's extracted automatically from its source code. The agent's source code itself is never sent to Collibra, only this extracted profile.
The SDK intercepts agents built with LangChain, LangGraph, and Deep Agents. You can add it to your agent in several ways, from two lines of code (activating Guardian and adding its middleware) to no code changes at all (launching the agent through the Collibra command line interface). The SDK needs a compiled bundle of the rules, which Guardian creates from the Agent Contract and the Agent Fingerprint. The bundle is downloaded with the Collibra command line interface. Without a bundle, Guardian allows every action and logs a warning.
Agent Fingerprint
An Agent Fingerprint describes one state of an AI agent, as found by analyzing its source code. The agent is never run, so the fingerprint describes the agent as written, not as observed. It contains:
- The agent and its subagents, and the tools each of them can call, including each tool's arguments, return type, and the checks that the tool's own code already makes.
- The system prompts and models of each agent, and the middleware that wraps it.
- Information about how the fingerprint was built, for example the platform and, optionally, the source repository and commit.
- Warnings about anything that couldn't be resolved. A list that couldn't be resolved means "unknown", not "none".
When something changes in the agent's tools, prompts, or models, a new fingerprint is created, and Guardian recompiles the rules from it.
Note TBC - the SDK is built for LangChain, LangGraph, and Deep Agents, and the Guardian code mentions Databricks as a platform. The Gateway proxy path wasn't found as built in the Guardian code (Groundwork, 2026-10-22), where it's described as a planned second host that reuses the same engine. Confirm with dev whether the Gateway is available in this private preview, and what runtime enforcement supports.
- For information on enabling Guardian Agents and Agent Contracts, go to Set up AI Command Center.
- Learn more about the Agent Contract model.
- Learn more about Guardian Agent runtime enforcement.