Diagrams

The agent creates architecture diagrams the way an experienced architect would. It thinks about layout, data flow direction, logical groupings, and visual hierarchy before placing anything. It uses tools to search for the right icons, compute optimal positioning, and visually inspect its own work to ensure the diagram is clear and well-organized.

How the agent builds diagrams

The agent follows a structured process. It plans the layout by analyzing connection topology, data flow, tiers, and groupings. It searches the icon library to find the right visual for every resource. It builds the diagram inside-out, starting with the innermost resources and working outward to containers and groupings. For diagrams with three or more nodes, it uses a layout algorithm to compute initial positions that avoid overlaps and minimize edge crossings.

After creating a diagram, the agent reviews its own work using a scoring system and makes improvements until the diagram meets quality targets.

Scoring system

Every diagram is evaluated against a set of visual quality metrics.

• Edge straightness measures whether connections are aligned cleanly
• Edge crossings penalizes lines that cross over each other
• Edge-node overlap checks if connections pass through unrelated resources
• Container correctness validates that groupings properly contain their children
• Spacing consistency ensures even gaps between resources
• Node overlap detects resources that sit on top of each other
• Composition evaluates the overall logical organization

The score adjusts for complexity. Simple diagrams are held to a higher standard than large, complex ones. You can see the score and use it as a guide to improve the diagram yourself.

Human-agent collaboration

Diagrams are not just agent-generated. You can drag and move resources and containers yourself. As you drag, alignment guides appear to help you snap elements into clean positions. Containers automatically resize to fit their children as you rearrange things.

This makes diagrams a great way to learn architecture design. Move things around, see how the score changes, and develop a feel for what makes a well-organized architecture diagram.

Visual edits mode

Press V or use the visual edits toggle in the chat input area to enter visual edits mode. In this mode, clicking on any resource or connection opens a contextual chat input where you can tell the agent what to change about that specific element. Click on empty space to make diagram-wide requests.

You can also right-click on any element to open the same contextual chat. This lets you work with specific parts of the diagram without describing which element you're referring to.

Multiple diagrams

You can create multiple diagrams per workspace and toggle between them from the diagram panel. This lets you maintain different views of the same infrastructure. For example, you might have an L1 high-level overview, an L2 service-level diagram, an L3 detailed component diagram, a physical cloud architecture view, or a threat vector map all in the same workspace.

Diagram types

• Architecture diagrams at different levels of detail showing how services, networks, and resources connect
• Physical cloud architecture showing the actual cloud provider layout
• Threat vector maps visualizing potential attack surfaces and security boundaries
• Flow diagrams illustrating data flow or request paths
• Dependency graphs mapping relationships between modules or services