How omi-kalix/multi-agent-ai-workflow-for-content-creation works
1. Repository purpose
This repo is not a traditional application codebase; it is a workflow-as-code package for a multi-agent content creation system. The main deliverable is an n8n workflow definition that automates LinkedIn post creation end to end.
2. Primary runtime boundary
The core execution environment is n8n. The repository stores the workflow configuration, while execution happens inside n8n’s orchestrator. That means the repo defines the automation graph, but the actual runtime depends on n8n plus external APIs and credentials.
3. Main orchestration flow
The system follows a linear content pipeline:
form input → research → content generation → image creation → email delivery.
This is the central data flow to represent in an architecture diagram.
4. Workflow subsystem
The file workflows/Multi-Agent-AI-Workflow-for-Content-Creation.json is the key artifact. It likely contains the node graph, branching logic, and service integrations that implement the multi-agent sequence. This is the “control plane” of the repo.
5. Prompt subsystem
prompts/prompts.md is the prompt library for the AI agents. Architecturally, this is the layer that separates agent behavior/specification from the workflow wiring. The prompts define how research, writing, and possibly packaging behave without embedding that logic directly in the workflow definition.
6. Research boundary
The research step is an external data acquisition stage using Crawleo.dev as the search/research service. This is the first place where live web context enters the system, and it feeds structured input into the downstream content-generation agent.
7. AI generation boundary
The content-writing agent uses a local or self-hosted model stack centered on Ollama / Llama 3.2 per the README. This is the main text-generation boundary and represents the system’s reasoning/content synthesis layer.
8. Visual generation boundary
The visual agent produces a custom image using external image-generation services such as DALL·E or Imagen 4. In architecture terms, this is a separate modality pipeline: it consumes the topic/context and outputs a post image asset.
9. Delivery subsystem
The final step packages the generated post and image and sends them through an email service such as Gmail or SendGrid. This is the output boundary of the automation, turning workflow results into a deliverable artifact for the user.
10. Assets directory
assets/ contains screenshots and example outputs, including workflow diagrams and generated post images. These are documentation artifacts only; they support understanding and validation but are not part of the runtime path.
11. Operational dependencies
The README indicates the system depends on API keys and credentials for AI, web search, and email. These external secrets are a key architectural dependency because they gate access to each stage of the pipeline.
12. Deployment/usage model
The repo is designed for import into an n8n environment rather than being built and deployed like a software service. The practical deployment unit is the workflow import plus configuration of connected services.
13. What to emphasize in a diagram
A clear diagram should show:
the user submitting a topic into n8n, n8n coordinating separate research/content/image/delivery stages, and each stage calling an external service. The most important boundaries are n8n orchestration, research API, LLM generation, image generation, and email delivery.
14. Overall architecture summary
This is a multi-agent, service-oriented automation workflow with a single orchestrator and several specialized external providers. The repo’s job is to capture that orchestration and the prompts that shape the agents, not to implement custom application logic.
Main entry points: prompts.md Prompt Library · Multi-Agent-AI-Workflow-for-Content-Creation.json Workflow JSON
Components 14
Repo Artifacts
Workflow package contents and supporting documentation.
- Workflow JSON · Main n8n workflow definition for the content-creation pipeline.
workflows/Multi-Agent-AI-Workflow-for-Content-Creation.json - Prompt Library · Agent instructions and behavior spec separated from workflow wiring.
prompts/prompts.md - Assets · Screenshots and sample outputs used for documentation and validation.
assets - API Keys · Operational credentials required for research, generation, and delivery.
n8n Runtime
The orchestrated automation flow that executes inside n8n.
- n8n · Execution boundary that runs the workflow graph and coordinates all stages.
- Topic Intake · Starting point where the user submits a topic or content request.
- Research Agent · Collects web context and structured notes for downstream generation.
- Writing Agent · Synthesizes the LinkedIn post content from topic and research inputs.
- Image Agent · Creates a custom visual asset for the post.
- Delivery Agent · Packages the generated post and image for email delivery.
External Services
Third-party APIs and model providers used by the workflow.
- Crawleo.dev · External research service used to gather live web context.
- Ollama / Llama 3.2 · Local or self-hosted model stack used for post generation.
- DALL·E / Imagen 4 · External image-generation service used for the visual asset.
- Gmail / SendGrid · Outbound delivery service for sending the finished package.
Connections 16
- Workflow JSON → n8n: imports · Workflow definition is imported into the n8n execution environment.
- n8n → Topic Intake: starts · n8n receives the submitted topic and begins orchestration.
- Topic Intake → Research Agent: feeds · Topic input is handed to the research stage.
- Research Agent → Crawleo.dev: queries · Research stage pulls external context from Crawleo.dev.
- Research Agent → Writing Agent: context · Structured research notes are passed into the writing agent.
- Prompt Library → Writing Agent: instructs · Prompt library shapes post style, structure, and behavior.
- Writing Agent → Ollama / Llama 3.2: generates · Writing agent calls the LLM runtime for content synthesis.
- Writing Agent → Image Agent: handoff · Generated topic and copy context are forwarded to the visual stage.
- Prompt Library → Image Agent: guides · Prompt library can also define image-generation behavior.
- Image Agent → DALL·E / Imagen 4: renders · Image stage requests the external image provider.
- Image Agent → Delivery Agent: asset · Generated image is packaged with the post for delivery.
- Delivery Agent → Gmail / SendGrid: sends · Final package is delivered through the email service.
- API Keys → Crawleo.dev: auth · API credentials gate access to the research service.
- API Keys → Ollama / Llama 3.2: auth · API credentials or local access gate the model runtime.
- API Keys → DALL·E / Imagen 4: auth · API credentials gate access to image generation.
- API Keys → Gmail / SendGrid: auth · API credentials gate outbound email delivery.