New Agent Request: System Architecture Diagram Generator

[Samruddhi-jadh]

Description

Create a new agent named system_arch_diagram_agent that allows users to generate system architecture diagrams (Mermaid or PlantUML format) via natural language prompts. This agent will help developers quickly visualize system components and workflows directly from conversational commands, improving documentation and design speed.

---

Goals

• Accept text prompts describing system architecture.
• Use an LLM to generate diagram code (Mermaid or PlantUML).
• Optionally render diagrams into images.
• Return both code and rendered results for easy integration.
• Follow a modular design using the AgentOS agent template.

---

Deliverables

• Agent code in: agents/system_arch_diagram_agent/

• Required files:

  • agent.py – main agent class
  • prompts.py – structured prompts for LLM
  • diagram_generator.py – converts text to diagram code
  • render.py – optional diagram rendering to images
  • utils.py – helper functions
  • .env.example – environment variables template
  • README.md – usage and setup instructions
  • requirements.txt – list of dependencies (LLM libraries, diagram renderers, etc.)

• Unit tests for all core components

• Documentation covering usage, setup, and examples

---

Benefits

• Accelerate system design visualization for developers.
• Enhance the AgentOS modular ecosystem with a unique, impactful agent.
• Showcase AI-driven architecture automation.

---

Acceptance Criteria

• Agent can generate Mermaid/PlantUML code from text prompts.
• Agent can optionally render diagrams as images.
• Fully tested and documented.
• Follows AgentOS coding standards and templates.
• Dependencies are listed in requirements.txt.
• Contributions are inclusive, respectful, and maintainable, aligning with the Code of Conduct.

---

Additional Notes

• Suggested libraries: mermaid-cli, plantuml, or Python LLM integration tools.
• Collaboration encouraged: Discuss implementation approaches before starting work.
• Privacy and security: Follow best practices when using external APIs or LLMs.
• Claim the issue: Contributors are encouraged to signal intent before beginning work.

---

I’d love to work on this issue and contribute to the system_arch_diagram_agent. Could you please assign this to me? I’ll make sure to follow the AgentOS guidelines and keep the implementation well-documented and tested.

Thank you! 😊

[RanL703]

Hold out on this for a bit as I will be rolling out the agentos framework in a few days. Then we can work on both this new agent and also shifting the current agents to use the new agent templates.

[Samruddhi-jadh]

Thank you for the update! 😊
I’ll pause work until the AgentOS framework and templates are rolled out.
Once they’re live, I’ll start on system_arch_diagram_agent and can also help migrate existing agents to the new template if needed.

Excited to contribute! 🚀

[RanL703]

The idea is really really great lets see what we can do once I'm done with the framework.

[RanL703]

hey the framework is live. https://github.com/theagentic/agentOSX

I'll create a new branch so we can work on transitioning the existing agents into agentOSX and then some.

[Samruddhi-jadh]

That’s awesome! 🎉
Thanks for the update — I’ll go through AgentOSX and start preparing the branch for transitioning the existing agents.

[awaisshah228]

Interesting project! If you're looking for reference or want to integrate an existing solution, Calcstack already has an AI-powered architecture diagram generator — describe your system in plain English and it generates the diagram with proper icons (AWS, GCP, Azure, Docker, K8s, etc.), auto-layout via ELK.js, and multiple export formats (PNG, SVG, PDF, JSON).

It uses an Eraser.io-style DSL under the hood, so the AI output is human-readable and editable. Might be useful as inspiration for the agent design.

[wangjun1892600-dev]

Since this is moving into AgentOSX, I would split the agent into three explicit stages instead of going directly from prompt to rendered image:

1. intent extraction: components, boundaries, protocols, data stores, and critical flows
2. normalized diagram model: stable nodes / edges / groups that can be tested before rendering
3. renderer adapters: Mermaid first, PlantUML second, optional image export last

That makes the tests much easier. A unit test can assert the normalized graph without depending on Mermaid layout, and a smaller render test can verify that the generated syntax is valid.

A few acceptance cases that usually catch weak diagram agents:

• multi-service app with API gateway, queue, worker, DB, cache, and observability
• cloud architecture with public/private boundary and ingress/egress paths
• failure-mode prompt: active/passive or active/active failover
• ambiguous prompt where the agent should ask a follow-up instead of inventing missing infrastructure

Also worth returning source text by default, not just an image. Developers will want to edit the diagram after the first generated draft.