Agentic Research — Shared Team Playbook

This is the shared playbook loaded by ALL agents in the agentic research system. It defines the protocols, quality standards, and conventions that ensure consistent, high-quality research across the orchestrator and all specialist agents.

Every agent MUST read and follow this playbook before beginning work.

Team Structure

Agent Role Key Tools

Orchestrator Coordinates workflow, relays URLs, synthesizes report TeamCreate, SendMessage, Task

ExaAI Specialist Semantic web search, source discovery mcp__exa__web_search_exa

Firecrawl Specialist Web scraping, site crawling, vector DB firecrawl CLI, Qdrant

NotebookLM Specialist AI-powered analysis, artifact generation notebooklm CLI

Communication Protocol

Message Format

All inter-agent messages MUST use this structure:

[Agent Name] [Message Type] (Batch N)

[Section 1 Header]

• Item 1
• Item 2

[Section 2 Header]

• Item 1

Summary

Brief summary of key points.

Message Types

Type When From To

URL Report Every 3-5 operations Specialists Orchestrator

Progress Update Every 5-10 operations Specialists Orchestrator

Crawl Request When docs site discovered Orchestrator Firecrawl

Source Relay When high-quality URLs received Orchestrator NotebookLM

Completion Signal When all work done Specialists Orchestrator

Error Escalation When blocked Any agent Orchestrator

URL Report Format (Specialists -> Orchestrator)

[Agent] URL Report (Batch N)

Key URLs Discovered:

• https://url1.com - [Description] - Tier: [Primary/Academic/Official/Industry/Community]
• https://url2.com - [Description] - Tier: [tier]

Documentation Sites (recommend for Firecrawl):

• https://docs.example.com - Full documentation for X

Key Findings:

• Finding 1
• Finding 2

Source Relay Format (Orchestrator -> NotebookLM)

Add these N sources to the notebook:

• https://url1.com — [Description] — Tier: Primary
• https://url2.com — [Description] — Tier: Academic

Completion Signal Format

[Agent] specialist complete. Findings written to findings/[agent]-findings.md. [Key stats: searches, URLs, pages, etc.] Key gaps: [list gaps].

Source Quality Tiers

All agents MUST classify discovered sources using these tiers. The orchestrator uses tiers to prioritize which URLs to relay to NotebookLM.

Tier Description Priority Examples

Primary Original research, official announcements, creator content Highest Blog posts by framework creators, official release notes, original research papers

Academic Peer-reviewed papers, university research, conference proceedings High arXiv papers, ACM/IEEE publications, NeurIPS/ICML proceedings

Official Official documentation, guides, tutorials from project maintainers High docs.langchain.com, readthedocs sites, official GitHub READMEs

Industry Analysis from recognized industry analysts, reputable tech publications Medium Hacker News (top posts), reputable tech blogs, benchmark reports

Community Forum discussions, personal blogs, Stack Overflow, Reddit Lower Reddit threads, personal dev blogs, community wikis

News News articles, press releases, general media coverage Lowest TechCrunch, The Verge, general news sites

Tier Prioritization Rules

• Primary + Academic: Always relay to NotebookLM

• Official: Always relay to NotebookLM (essential for technical accuracy)

• Industry: Relay if relevant to key questions and not redundant

• Community: Only relay if contains unique insights not found in higher tiers

• News: Rarely relay -- only if contains exclusive information

Source Limit Management

• NotebookLM limit: 300 sources (paid plan), 50 (standard)

• Target: 200 excellent sources > 300 mediocre ones

• Track count: Orchestrator maintains running total

• Stop early: If quality sources are exhausted before hitting limit

See references/source-quality-tiers.md for detailed tier classification examples.

Parallel Safety Rules

NotebookLM Commands

CRITICAL: ALL agents MUST follow these rules.

• ALWAYS use -n <notebook_id> or --notebook <notebook_id> flags

• NEVER use notebooklm use <id> -- it modifies shared state and is unsafe in parallel workflows

• Use full UUIDs in automation to avoid ambiguity from partial IDs

• Use --new flag on ask commands when switching topic areas (avoids conversation ID conflicts)

File System Safety

• Each specialist writes to its own findings file (no conflicts):

• ExaAI: findings/exa-findings.md

• Firecrawl: findings/firecrawl-findings.md

• NotebookLM: findings/notebooklm-findings.md

• Only the orchestrator writes to report.md and sources/sources.md

• Artifacts are downloaded to artifacts/ (NotebookLM specialist only)

Team Communication Safety

• All messages go through SendMessage tool (not shared files)

• Messages are automatically delivered -- no polling needed

• Each agent maintains its own state independently

• The orchestrator is the only agent that reads all findings files (during synthesis)

Output Formatting Standards

Findings Files

All specialist findings files MUST include:

• Research Topic -- exact topic from brief

• Methodology Summary -- what was done, key stats

• Key Findings -- organized by topic area, with inline source URLs

• Contradictions -- where sources disagree (do NOT smooth over)

• Gaps -- what couldn't be found or needs more investigation

Source Citation Format

Inline citations use markdown links:

According to the official documentation Source Title, the framework supports...

When citing multiple sources:

Multiple sources confirm this approach Source A, Source B, though Source C disagrees.

URL Reporting Format

When reporting URLs, always include:

• Full URL (no shortlinks)

• Brief description (what the page contains)

• Quality tier classification

• https://docs.example.com/guide - Comprehensive setup guide - Tier: Official

Error Handling Protocol

Severity Levels

Level Action Examples

Warning Log and continue Rate limit on one site, single source fails to load

Error Report to orchestrator, attempt workaround Auth failure, major site blocked, tool crash

Critical Report to orchestrator immediately, stop work All tools failing, auth completely broken

Error Reporting Format

Error Report

Agent: [name] Severity: [Warning/Error/Critical] Error: [description] Action Taken: [what you tried] Impact: [what's affected] Suggested Fix: [if known]

Recovery Strategies

• Rate limits: Exponential backoff (30s -> 2min -> skip)

• Auth failures: Run diagnostic commands, report to orchestrator

• Tool crashes: Retry once, then report

• Network issues: Wait 1 minute, retry, then report

• Source not found: Log, skip, continue with other sources

Persistent Memory Conventions

All agents write session learnings to ~/.claude/memory/ :

Agent Memory File

Orchestrator ~/.claude/memory/agentic-orchestrator.md

ExaAI ~/.claude/memory/exa-specialist.md

Firecrawl ~/.claude/memory/firecrawl-specialist.md

NotebookLM ~/.claude/memory/notebooklm-specialist.md

Memory Entry Format

Session: YYYY-MM-DD -- <topic-slug>

What worked:

• [insight 1]
• [insight 2]

What didn't work:

• [issue 1]

Lessons learned:

• [lesson 1]

Memory Usage

• Read memory at session start to leverage past learnings

• Write memory at session end with insights from current session

• Keep entries concise -- focus on actionable insights

• Update or remove entries that turn out to be wrong

Cross-References

This playbook is used by:

• Orchestrator agent: agents/agentic-orchestrator.md

• ExaAI specialist agent: agents/exa-specialist.md

• Firecrawl specialist agent: agents/firecrawl-specialist.md

• NotebookLM specialist agent: agents/notebooklm-specialist.md

Related skills:

• Orchestration methodology: skills/agentic-research-orchestration/SKILL.md

• ExaAI search: skills/exa/SKILL.md

• Firecrawl web scraping: skills/firecrawl/SKILL.md

• NotebookLM automation: skills/notebooklm/SKILL.md