Agentic Patterns

Agentic Design Patterns — draw.io Diagramming

Patterns from "Agentic Design Patterns" (Gulli & Sauco, 2025) applied to AI-powered diagram generation and multi-platform embedding

Applied Patterns

Prompt Chaining

Relevance: Generating a high-quality draw.io diagram is a multi-stage pipeline — understanding intent, selecting diagram type, producing XML structure, applying styles, validating output, and embedding. Each stage feeds the next. Current Implementation: Individual commands (drawio:create, drawio:style, drawio:export, drawio:embed) handle discrete steps. Users chain them manually. Enhancement: Formalise a diagram generation chain: (1) intent extraction → extract entities, relationships, and domain from user input; (2) type selection → route to the correct diagram type from the 196-type catalog; (3) skeleton generation → produce minimal valid XML structure; (4) style application → apply theme and professional defaults; (5) quality critique → invoke the quality-critique skill; (6) export/embed → route to target platforms. Each step's output is the next step's input. Failed steps trigger targeted regeneration, not full restarts.

Reflection

Relevance: AI-generated XML diagrams frequently contain layout collisions, missing connections, style inconsistencies, or structural errors that degrade diagram quality. Current Implementation: The quality-critique skill exists in skills/quality-critique/SKILL.md and provides post-generation analysis. Enhancement: Embed reflection as an automatic loop within diagram generation, not a separate manual step. After initial XML generation, a reflection agent scores the diagram on five axes — completeness, layout clarity, style consistency, semantic accuracy, and platform compatibility — and produces a structured critique. The generator incorporates the critique and regenerates targeted sections. Allow up to three reflection passes before delivering the final diagram. Surface the final quality score to the user.

Planning

Relevance: Complex diagrams (C4 architecture, network topology, BPMN processes) require upfront layout strategy to avoid crowded, unreadable output. Ad hoc generation produces poor spatial organisation. Current Implementation: The diagram-architect agent uses domain knowledge to guide structure, but planning is implicit and embedded in the prompt. Enhancement: Add an explicit planning phase before XML generation. The planning agent produces: (1) a layer plan (which semantic layers will be created); (2) a spatial grid (rough coordinates for major element groups); (3) a connection manifest (all edges with source, target, and label); (4) a style plan (colour theme, shape families, font sizes). The XML generator then fills in this plan rather than generating freely, producing more predictable and maintainable output.

Tool Use

Relevance: Diagram generation requires coordinated use of multiple tools: code analysis tools (Grep, Glob, Read) to discover architecture facts, XML validation tools, MCP diagram servers for live editing, and embedding tools for each target platform. Current Implementation: Commands invoke the appropriate tools within their scope. The MCP server integration (drawio:mcp-setup) enables programmatic diagram operations. Enhancement: Define a tool manifest for diagram generation that explicitly maps diagram phases to tool sets. Code-analysis tools are called during the intent-extraction phase; XML tools during generation and validation; MCP tools during live-editing sessions; platform-specific export tools during embedding. Agents should declare their tool intent before calling, enabling tool-use logging and rate limiting for expensive operations (e.g., Firecrawl calls during enrichment).

Routing

Relevance: The plugin supports 196 diagram types across 20+ domains. Selecting the wrong type for a given context produces a diagram that fails to communicate its intent. Current Implementation: The drawio:auto-diagram command uses heuristics to select diagram type. The CLAUDE.md routing table maps contexts to types. Enhancement: Implement a diagram-type routing agent that scores multiple candidate diagram types for the given input using weighted criteria: domain fit, entity count, relationship density, target audience, and platform constraints. Present the top-3 candidates with rationale before generating, allowing the user to confirm or redirect. Store routing decisions in a session log to improve future routing.

Parallelization

Relevance: Multi-format export (SVG, PNG, PDF, .drawio.svg) and multi-platform embedding (GitHub, Confluence, Azure DevOps, Notion, Teams, Harness) are independent operations that do not need to be sequential. Current Implementation: drawio:export and drawio:batch handle multiple outputs but execute sequentially. Enhancement: Parallelise export and embedding operations. After a diagram is finalised, spawn parallel export workers — one per output format — and parallel embedding workers — one per target platform. A merge agent collects results, surfaces any per-platform failures without blocking others, and produces a summary manifest of all exported artefacts with their paths and embedding snippets.

Resource-Aware

Relevance: Very large diagrams (200+ nodes, complex C4 systems, full enterprise architecture maps) can produce XML files that are expensive to process, slow to render in draw.io, and difficult to read without progressive disclosure. Current Implementation: No resource budgeting exists. Complex diagrams are generated at full fidelity regardless of output size. Enhancement: Implement resource-aware diagram generation with three complexity tiers: (1) standard — up to 50 nodes, full fidelity; (2) simplified — 51–150 nodes, group related elements into containers with expandable child diagrams; (3) summary — 151+ nodes, generate a high-level overview diagram with links to per-domain detail diagrams. Estimate output size from the planning phase and select the appropriate tier automatically. Warn the user when dropping to a lower tier and offer to generate detail pages on demand.

Pattern Interaction Map

User Intent / Source Code / Description
        │
        ▼
   [ROUTING] ──── select diagram type from 196-type catalog
        │
        ▼
   [PLANNING] ──── layer plan + spatial grid + connection manifest
        │
        ├─ Check entity count ──► [RESOURCE-AWARE] ──► tier selection
        │
        ▼
   [TOOL USE] ──── code analysis / MCP server / XML tools
        │
        ▼
   XML Generation
        │
        ▼
   [REFLECTION] ──── quality critique (up to 3 passes)
        │
        ▼
   Final Diagram XML
        │
        ▼
   [PARALLELIZATION] ──── concurrent export + platform embedding
        │
        ▼
   [PROMPT CHAINING] ties all stages together as a managed pipeline

Key interactions:

• Prompt Chaining is the meta-pattern that sequences all others into a coherent pipeline.
• Planning + Resource-Aware: The planning phase determines entity count, triggering tier selection before any XML is generated.
• Routing + Planning: Routing determines the diagram type, which constrains the planning agent's spatial and layer choices.
• Tool Use + Reflection: Reflection may trigger additional Tool Use calls (re-reading source code) to validate that the diagram accurately reflects the codebase.
• Parallelization is the terminal stage, running only after Reflection has approved the diagram quality.

References

• Gulli, A. & Sauco, M. (2025). Agentic Design Patterns. O'Reilly Media.
• Repository: github.com/Mathews-Tom/Agentic-Design-Patterns
• draw.io / diagrams.net: https://www.diagrams.net
• draw.io XML Reference: https://jgraph.github.io/mxgraph/docs/js-api/files/model/mxCell-js.html