lo6

Agentic Interface Specification

Overview

This document defines the LangGraph architecture for lo6. We utilize a stateful graph approach where agents (Nodes) modify a shared state object.

Shared Graph State

The NewsroomState interface defines the data passed between agents.

interface NewsroomState {
  // Messages for chat history
  messages: BaseMessage[];
  // The core entities (Optional as they populate over time)
  lead?: Lead;
  story?: Story;
  research?: ResearchDossier;
  assets?: Asset[];
  // Workflow control
  next_agent: string;
  human_feedback?: string;
  errors?: string[];
  // Status flags
  lead_status?: 'NEW' | 'APPROVED' | 'REJECTED' | 'MONITOR';
}

Workflows (Sub-Graphs)

1. Triage Workflow

Goal: Ingest raw data and filter for relevance.

• Nodes:
  • TriageManager: Orchestrator. Spawns specialized tools based on source type.
  • RSSFetcher: Parses RSS feeds.
  • APIConnector: Connects to external APIs.
  • WebCrawler: Scrapes and crawls websites.
  • SignalReceiver: Listens for verified leads from the lo6 Network.
  • SignalReceiver: Listens for verified leads from the lo6 Network.
  • ConfigAgent: Filters items based on “Interests”. Scores relevance (0-100).
  • HumanTriage: Interrupt Node. Waits for user to Approve/Reject/Note.
• Edges:
  • TriageManager -> RSSFetcher / APIConnector / WebCrawler / SignalReceiver
  • RSSFetcher / APIConnector / WebCrawler / SignalReceiver -> ConfigAgent
  • ConfigAgent -> (Score > 50?) -> HumanTriage
  • ConfigAgent -> (Score <= 50?) -> END

2. Editorial Workflow

Goal: Turn an approved Lead into a Draft Story.

• Nodes:
  • EditorialAgent: Orchestrator. Decides if research is needed.
  • ResearchAgent: Uses tools to gather facts.
  • JournalistAgent: Writes the markdown content.
  • ImageGenAgent: Generates assets based on story content.
  • HumanEditor: Interrupt Node. Review draft.
• Edges:
  • EditorialAgent -> ResearchAgent
  • ResearchAgent -> JournalistAgent
  • JournalistAgent -> ImageGenAgent
  • ImageGenAgent -> HumanEditor

3. Publishing Workflow

Goal: Distribute approved content.

• Nodes:
  • PublisherAgent: Formats content for specific platforms (Twitter thread vs Blog post).
  • HumanPublisher: Interrupt Node. Final “Go” button.
  • Distributor: API calls to external platforms.

Agent Tools Definition

Agent	Tools Available	Description
Config Agent	calculate_relevance(text, interests)	Local NLP scoring function.
Research Agent	tavily_search(query) scrape_webpage(url)	Web browsing capabilities.
Journalist Agent	read_research_dossier() check_style_guide()	Context retrieval tools.
Image Gen Agent	dalle_3(prompt) stable_diffusion(prompt)	Image generation APIs.
Publisher Agent	twitter_api_post() wordpress_api_post()	External publishing tools.

System Prompts Strategy

• Role: “You are an expert [Role Name]…”
• Context: “You are working on a story about [Title]…”
• Constraint: “Output must be valid JSON/Markdown…”
• Security: “Do not invent facts. Cite sources from the Research Dossier.”

This site is open source. Improve this page.