Map each AI system. Know its blast radius.

1. 1

   Create an AI system for one job your organization runs.

2. 2

   Attach the agents, applications, MCP servers and models Praesidia already governs, and give each one its role in the system.

3. 3

   Record how assets depend on one another as directed relationships, from your own tooling through the API.

4. 4

   Open the system’s Graph tab, expand assets one hop at a time, and filter by type, environment, owner or data sensitivity.

5. 5

   Select an asset to see which AI systems depend on it, and highlight the path each one takes.

AI systems and their assets

An AI system groups the assets it is built from — such as applications, agents, models and model endpoints, MCP servers and their tools, APIs, data sources, datasets, vector stores, RAG indexes, prompts, skills, vendors, identities, credentials, repositories, cloud resources and workflows — and records each asset's role in it: primary, supporting, dependency or external. The same asset can belong to more than one AI system.

Adopting what you already govern

Agents, applications, MCP servers and model configurations that Praesidia already governs can be adopted as assets and attached to an AI system in one step, with their role; through the API, workflows and evaluation datasets can be adopted too. Adopting something that is already an asset returns the existing asset instead of creating a duplicate.

Directed relationships

Relationships between assets are directed and say what they mean: uses, calls, accesses, reads, writes, contains, delegates to, hosted by, has permission, governed by, can invoke, grants scope and can assume. An agent that calls a server and a server that calls an agent are different relationships, so what an asset depends on and what depends on it are separate questions with separate answers. Relationships can be recorded, edited, archived and restored through the API.

Multi-hop queries

One query through the API walks the graph from any asset for several hops — downstream to what it uses, upstream to what depends on it, or both — three hops by default and up to six. The walk can be narrowed to particular asset types, relationship types or one data-residency region, and archived items stay out unless you ask for them.

The graph view

Every AI system has a Graph tab: the system at the centre, its assets around it labelled with their role and type, and each relationship drawn as an arrow labelled with its type. Expand an asset to bring in its neighbours one hop at a time, filter by asset type, environment, owner or data sensitivity, and select an asset to see its type, owner, environment, source and data classification. Tab to an asset and press Enter to inspect it from the keyboard.

Blast radius

Select an asset on the graph and Praesidia lists every AI system that depends on it within the hop limit — its blast radius — most critical first, with the shortest path and the highest-risk path from each system to that asset. Either path can be highlighted on the graph, and the same answer is available through the API.

Where it stops

The graph shows the AI systems, assets and relationships recorded in your organization; a dependency nobody has recorded is not on it. A query that asks for more hops than the cap is answered to the cap and says so, and a query too large to answer in full is refused rather than returned in part.

Read more: agent topology maps and dependency graphs · building an AI agent inventory. Then: agent discovery · risk management · assurance · MCP server governance.

The graph and its examples on this page are illustrative. New policies start in observe mode: decisions are recorded and nothing is blocked until you switch a policy to enforce. Approving a held call is in every plan; multi-step approval workflows are Enterprise; sequence rules and red-teaming are on Advanced.