Agentic Research Orchestration Skill

YOU MUST invoke this skill (NOT optional) when the user mentions ANY of these triggers:

• "deep research", "comprehensive research", "agentic research"

• "multi-source research", "research with multiple agents"

• "investigate thoroughly", "thorough analysis with sources"

• Any mention of /agentic-research or coordinated research workflows

Failure to invoke this skill when triggers occur violates your operational requirements.

This skill defines the orchestrator's methodology for conducting deep agentic research. Only the orchestrator agent loads this skill. Specialists load the shared agentic-research skill plus their domain-specific skills.

Overview

As the orchestrator, you manage a 5-phase research workflow:

• Clarification -- Gather complete requirements from user

• Setup -- Create infrastructure (directories, research brief, NotebookLM notebook)

• Dispatch -- Spawn and coordinate 3 specialist agents

• Active Orchestration -- Relay URLs, cross-pollinate discoveries, monitor progress

• Synthesis -- Generate artifacts, write final report, notify user, shutdown team

Phase 1: Clarification Protocol

Parse the research topic from your initial prompt. Then use the AskUserQuestion tool to ask a minimum of 5 clarifying questions before proceeding.

DO NOT skip this phase. DO NOT make assumptions about scope or depth.

Required Questions

Ask ALL of these questions, plus any additional ones needed for zero ambiguity:

• Scope boundaries -- What is explicitly in scope? What is explicitly out of scope? Are there adjacent topics to include or exclude?

• Depth requirements -- Do you need a surface-level overview, a mid-depth analysis, or a deep technical dive with primary sources and citations?

• Target audience -- Who will read this? (developer, executive, academic researcher, general audience, other)

• Desired output format -- What form should the final deliverable take? Options: strategic report, technical analysis, comparison matrix, literature review, decision brief, other.

• Key questions that MUST be answered -- List the 3-5 specific questions that this research must definitively answer.

• Known sources or starting points -- Are there specific URLs, papers, tools, or authors the user already knows about?

• Time sensitivity -- Should this focus on cutting-edge 2025-2026 developments, historical context, or both?

Clarification Guidelines

• Continue asking follow-up questions until there is ZERO ambiguity about what the user wants

• If the user gives a vague answer, probe deeper

• If conflicting requirements emerge, ask the user to prioritize

• The quality of the entire research operation depends on this phase

Document the Research Brief

Once all questions are answered, you will write a research brief in Phase 2. For now, keep detailed notes about: exact scope boundaries, depth level, target audience, output format, key questions, known sources, and time sensitivity.

Phase 2: Setup

Create Output Directory

mkdir -p ./docs/research/YYYY-MM-DD-<topic-slug>/findings mkdir -p ./docs/research/YYYY-MM-DD-<topic-slug>/sources mkdir -p ./docs/research/YYYY-MM-DD-<topic-slug>/artifacts

Write Research Brief

Write ./docs/research/YYYY-MM-DD-<topic-slug>/research-brief.md with: topic, scope, depth, target audience, output format, key questions, known sources, time sensitivity, and additional context.

Create NotebookLM Notebook

notebooklm create "Agentic Research: <topic>" --json

Extract and store the notebook ID. For ALL subsequent NotebookLM commands, use the -n <notebook_id> flag. NEVER use notebooklm use -- that command is not safe for parallel multi-agent execution.

Phase 3: Dispatch Specialist Agents

Create Team

Use TeamCreate with:

• team_name : agentic-research-<topic-slug>

• description : Deep research on <topic>

• agent_type : agentic-orchestrator

Spawn Specialist Agents

Spawn exactly 3 specialist agents via the Task tool with the team_name parameter:

Agent subagent_type Receives Special Notes

ExaAI exa-specialist

Research brief, output directory Conducts 10-20 semantic searches

Firecrawl firecrawl-specialist

Research brief, output directory Maps/crawls sites, auto-embeds to Qdrant

NotebookLM notebooklm-specialist

Research brief, notebook ID, output directory Always use -n <notebook_id> flag

Phase 4: Active Orchestration Loop

This is the CRITICAL coordination phase. You are the hub. You must actively manage information flow.

URL Relay Protocol

• Monitor incoming messages from specialists

• When specialists report discovered URLs:

• Evaluate each URL for relevance

• Maintain a running count (max 300 sources for NotebookLM)

• Cherry-pick the BEST URLs: primary > secondary, academic/official > blogs, sources answering key questions, unique perspectives, recent if time-sensitive

• Relay selected URLs to NotebookLM specialist via SendMessage

• When a documentation site is discovered, message the Firecrawl specialist to map and crawl it

Orchestration Rules

• Do NOT wait passively -- check if specialists need input when idle

• Do NOT send duplicate URLs to NotebookLM

• Prioritize quality over quantity for NotebookLM sources

• Cross-pollinate discoveries between agents

• Keep orchestrating until ALL 3 specialists signal completion

Phase 4.5: Artifact Generation

Once all specialists have reported findings are complete, request artifact generation from the NotebookLM specialist (report, mind map, data table). Wait for artifact completion before proceeding.

Phase 5: Synthesis and Completion

• Read all findings files from all specialists

• Query Qdrant for gaps in coverage

• Write final report at <output_dir>/report.md

• Write deduplicated sources file at <output_dir>/sources/sources.md

• Send Gotify notification (MANDATORY)

• Write to persistent memory

• Shutdown team (send shutdown_request to each specialist, then TeamDelete)

• Present results to user

Important Implementation Notes

• Parallel Safety: Always use -n <notebook_id> with notebooklm commands

• NotebookLM Source Limit: Maximum 300 sources per notebook (paid plan)

• Firecrawl Auto-Embedding: All scrape/crawl operations auto-embed to Qdrant

• Quality over Speed: Deep research takes 30-60 minutes -- do not rush

Cross-References

• Agent spawn patterns: references/agent-spawn-patterns.md

• Output templates: references/templates.md

• Orchestration transcript example: examples/orchestration-transcript.md

Agent Tool Usage Requirements

CRITICAL: When invoking scripts from this skill via the zsh-tool, ALWAYS use pty: true .