{“content”:”---\nname: organizational-agent-newsroom\ndescription: CrewAI/AutoGen-style multi-agent newsroom using Hermes delegate_task. CEO coordinates parallel Scouts, Critic fact-checker, and Editor synthesizer for morning briefings.\n---\n\n# Organizational Agent Newsroom — Hermes Native Implementation\n\n## Concept & Vision\n\nMimic CrewAI/AutoGen-style multi-agent orchestration using only Hermes’s native delegate_task. \nA “CEO agent” coordinates parallel “Scout” researchers, a “Critic” fact-checker, and an “Editor” synthesizer — \nproducing structured morning news briefings with zero hallucination tolerance.\n\nKey insight: Paperclip framework doesn’t exist yet. Use delegate_task with isolated subagents instead.\nEach subagent gets its own conversation, terminal session, and toolset. Results are returned as an array.\n\n## Architecture\n\n\nCEO (Jakephone — primary agent)\n├── Parallel Scouts (delegate_task, tasks=[])\n│ ├── Scout-A: AI / Technology\n│ ├── Scout-B: Malaysian Economy\n│ ├── Scout-C: Interior Design Trends\n│ └── Scout-D: International Geopolitics\n├── Critic (delegate_task, isolated)\n│ └── Role: fact-checking, cross-validation, \"待验证\" tagging\n└── Editor (delegate_task, isolated)\n └── Role: synthesize final draft\n\n\n## Workflow (Morning Briefing, 7:35 AM)\n\n\n7:35 ──┬── CEO broadcasts task via delegate_task (parallel scouts)\n │\n7:35-7:40 ──┬── Scout A (parallel)\n ├── Scout B (parallel)\n ├── Scout C (parallel)\n └── Scout D (parallel)\n │\n7:40 ───────────┼── Four raw reports returned to CEO\n │\n7:40-7:43 ───────┼── Critic reviews all four\n │ • Verify facts, flag \"待验证\"\n │ • Check logical consistency\n │ • Reject fabrication\n │\n7:43-7:47 ───────┼── Editor synthesizes\n │\n7:47 ───────────┴── CEO final QA → deliver to user\n\n\n## Scout Prompt Template (per domain)\n\nEach Scout receives:\n\n\nRole: {domain} Scout for 全能新闻实验室\nTask: Research and summarize {domain_count} news items from the past 24 hours.\n\nDomain focus:\n{domain_description}\n\nHard rules:\n- Only report facts with data support or reliable sources\n- Tag uncertain items as \"待验证\"\n- No marketing hype, no speculation presented as fact\n- Max 3 key points per news item\n\nOutput format:\n## {Domain} — YYYY-MM-DD\n\n### [News Item 1]\n- 核心事实: ...\n- 影响评估: ...\n- 来源: [URL or source name]\n- 状态: 已核实 | 待验证\n\n### [News Item 2]\n...\n\nDeliver your final report to the CEO when complete.\n\n\n## Critic Prompt Template\n\n\nRole: Critic (事实审查员) for 全能新闻实验室\n\nYou will receive 4 scout reports. Your job:\n1. Cross-verify facts against independent sources (use web search)\n2. Flag any claim that sounds \"too perfect\" as \"待验证\"\n3. Check logical consistency across all 4 reports\n4. Reject any fabrication — send back to scout for correction\n\nFor each disputed claim:\n- State the original claim\n- Provide counter-evidence or confirmation\n- Recommend: 接受 / 待核实 / 拒绝\n\nIf all reports pass scrutiny, respond: \"APPROVED — proceed to Editor\"\nIf any report needs revision: \"REVISION REQUIRED — [scout name] needs to fix [item]\"\n\n\n## Editor Prompt Template\n\n\nRole: Editor (主编) for 全能新闻实验室\n\nYou will receive 4 verified scout reports + critic verdict.\n\nTask: Synthesize into a final morning briefing for Sozo.\n\nFormat requirements:\n- Total length: 800-1200 characters (reading time ~5-8 min as podcast)\n- Structure per section:\n 1. AI/科技前沿 (3 items max)\n 2. 国际政经热点 (3 items max)\n 3. 室内设计趋势 (2-3 items)\n 4. 马来西亚经济 (3 items, emphasize JS-SEZ, NETR, MYR)\n\nEach item format:\n**【事实】** 核心叙述\n**【影响】** 1-2句影响评估\n**【来源】** source\n\nStyle:\n- Cold, objective narration\n- Minimal adjectives, maximum data\n- Extremely restrained humor only if news is light\n- If heavy news day: pure rationality\n\nFinal output prepend: \"【全能新闻实验室 | YYYY-MM-DD 晨间简报】\"\n\n\n## Implementation (Hermes/Cron)\n\nbash\n# Cron job at 7:25 AM daily\n# Skills: daily-news-podcast-pipeline, organizational-agent-newsroom\n\n# CEO task: delegate 4 scouts in parallel\n# Wait for results, then delegate critic\n# Wait for critic approval, then delegate editor\n# CEO synthesizes and sends final to Sozo via Telegram\n\n\n## Pitfalls\n\n1. Subagents have no memory of main conversation — pass ALL context via context field\n2. Subagents CANNOT call: delegate_task, clarify, memory, send_message, execute_code\n3. Timeout: set max_iterations appropriately (30-50 for research tasks)\n4. PC CPU already at 100% — consider limiting concurrent subagents if PC is primary worker\n5. If critic rejects: loop back to specific scout for revision, then re-critic\n\n## Status\n\n- Paperclip not installed — using Hermes native delegate_task as workaround\n- Organizational pattern proven: CEO→Scouts→Critic→Editor→CEO is a valid workflow\n- This pattern is reusable for: research reports, competitive analysis, multi-domain summaries\n”}