Multi-agent content pipeline with human-in-the-loop validation. Planner → Researcher → Writer → Editor → Human Checkpoint → Publisher.

Built with LangGraph · LangChain · Streamlit · LinkedIn API · DuckDuckGo / Tavily

• Multi-Agent Orchestration: Planner → Researcher → Writer → Editor → Publisher
• Intelligent Revision Loop: Automatic rewriting when content scores below quality threshold
• Human Checkpoint: Pause execution for review, approval, or custom edits before publishing
• Dataset Export (JSONL): Automatically save approved content with full context (metadata, research, scores, feedback) for LLM training and RAG pipelines
• Multi-Provider LLM Support: DeepSeek (default), OpenAI, Anthropic, Google Gemini (free tier)
• Web Research Integration: DuckDuckGo (free) and Tavily (premium) search providers
• Platform-Aware Content: Optimize for LinkedIn, Twitter, YouTube with platform-specific formatting
• Run Resumption: SQLite-backed checkpointing — resume interrupted runs without losing progress
• Dual Interface: CLI for scripting, Streamlit UI for interactive review

• Orchestration: LangGraph (state machine-based agentic workflows)
• LLM Integration: LangChain with multi-provider support
• Persistence: SQLite with LangGraph checkpointer
• UI: Streamlit (human-in-the-loop interface)
• Monitoring: Loguru (structured logging)

1. Planner Agent — topic → specific content brief (angle, audience, key points)
2. Research Agent — web search, trends, data synthesis
3. Writer Agent — draft adapted to platform and format
4. Editor + Fact-Checker — quality score (0–1), feedback, triggers rewrite if below threshold
5. Human Checkpoint — review, edit, approve, or reject with feedback
6. Publisher Agent — posts to LinkedIn (or dry-runs in dev mode)
7. Dataset Export — approved content saved to JSONL (training_data.jsonl) with full audit trail

# 1. Install
pip install -r requirements.txt

# 2. Configure
cp .env.example .env
# Edit .env with your LLM provider key

# 3. Run (CLI)
python -m src.main --topic "Why RAG systems fail in production" --platform linkedin

# Or run (UI)
streamlit run ui/app.py

python -m src.main \
  --topic "LangGraph patterns for AI agents" \
  --platform linkedin \
  --language en \
  --context "Target senior engineers. Focus on production patterns, not toy examples."

python -m src.main --thread-id <thread-id-from-previous-run>

streamlit run ui/app.py
# Opens at http://localhost:8501
# Tab 1: Start new content pipeline
# Tab 2: Review, edit, approve, or request revision
# Tab 3: View run history

ContentAgentOS/
├── src/
│   ├── core/
│   │   ├── config.py          # Settings & environment variables
│   │   └── state.py            # ContentState TypedDict + factory
│   ├── agents/
│   │   ├── llm.py              # LLM factory (4 providers)
│   │   ├── planner.py          # Topic → brief
│   │   ├── researcher.py       # Web search + synthesis
│   │   ├── writer.py           # Brief + research → draft
│   │   ├── editor.py           # Draft → quality score + feedback
│   │   └── publisher.py        # Draft → LinkedIn (+ dataset export)
│   ├── graph/
│   │   └── builder.py          # LangGraph orchestration
│   ├── tools/
│   │   ├── search.py           # Web search (DuckDuckGo / Tavily)
│   │   └── dataset_exporter.py # JSONL export for LLM training
│   └── main.py                 # CLI entrypoint
├── ui/
│   └── app.py                  # Streamlit UI
├── docs/
│   ├── ARCHITECTURE.md         # System design & patterns
│   ├── SETUP.md                # Installation & configuration
│   └── DATASET_EXPORT.md       # Dataset format & usage guide
├── datasets/                   # Generated training data (JSONL)
├── requirements.txt
├── .env.example
└── README.md

pytest tests/

tail -f /tmp/contentagentos.log

After human approval, content is automatically exported to datasets/training_data.jsonl with full context:

• Topic, angle, target audience, key points
• Research notes and sources
• Draft versions and editor scores
• Human feedback and final edits
• Publishing status

Perfect for:

• LLM Fine-tuning: Train custom models on your approved content
• RAG Pipelines: Use approved content as reference material
• Quality Analysis: Track trends and scoring patterns
• Compliance Audits: Maintain audit trail of approvals

See DATASET_EXPORT.md for usage examples.

See SETUP.md for detailed configuration options.

LLM_PROVIDER: Which LLM to use (deepseek | openai | anthropic | gemini)
QUALITY_THRESHOLD: Auto-rewrite threshold (0.0–1.0, default 0.7)
MAX_RETRIES: Max automatic rewrite attempts (default 2)
DRY_RUN: Log instead of publish to LinkedIn (default true)

• ARCHITECTURE.md — System design, module responsibilities, data flow, patterns
• SETUP.md — Installation, configuration, troubleshooting, running the pipeline
• DATASET_EXPORT.md — Dataset format (JSONL), schema, usage examples, fine-tuning guide

• Batch content generation from CSV
• More platform integrations (Medium, Substack, Dev.to)
• Prompt optimization via LangSmith
• Evaluation framework (benchmark against competitors)
• Async agent execution for parallel processing
• Database persistence (runs analytics, trending topics)