collaborative-knowledge-compounding

Collaborative Knowledge Compounding

Pattern: Multi-agent iterative research with shared notebook workspace Based on: AgentRxiv (arXiv:2503.18102) - "Collaborative Autonomous Research" Key Finding: Agents with shared research repository achieve 11-14% better results through compounding knowledge

---

Usage

/collaborative-knowledge-compounding "Research topic" [num_agents] [num_iterations]

Arguments:

• $ARGUMENTS (required): Research topic or question
• num_agents (optional): Number of research agents (default: 3, max: 5)
• num_iterations (optional): Research rounds (default: 3, max: 7)

---

Architecture

Shared Workspace: Srcbook Notebook

All agents collaborate through a shared .src.md notebook that tracks:

• Research findings from each agent
• Hypotheses generated and tested
• Evidence collected
• Synthesis across iterations
• State tracker (deterministic)

Agent Specializations

Agent 1: Literature Scout

• Role: Find and retrieve academic papers
• Tools: arxiv-paper-mcp, exa (discovery), firecrawl (known URLs)
• Output: Raw paper data + metadata

Agent 2: Technical Analyst

• Role: Find implementation examples and code
• Tools: exa__get_code_context_exa, context7, firecrawl
• Output: Code examples + technical specs

Agent 3: Synthesis Specialist

• Role: Read other agents' findings, identify patterns
• Tools: Read notebook cells, compare findings
• Output: Cross-agent insights + research gaps

Agent 4: Hypothesis Generator (if num_agents > 3)

• Role: Propose testable claims based on synthesis
• Tools: Reads all prior work
• Output: Novel hypotheses to explore

Agent 5: Validation Specialist (if num_agents = 5)

• Role: Cross-check claims across agents
• Tools: Targeted searches to verify/refute
• Output: Confidence scores + contradictions

---

Workflow

Phase 0: Initialize Shared Notebook

Creates: .collaborative-research/research-notebook.src.md

<!-- srcbook:{"language":"typescript"} -->

# Collaborative Research: [Topic]
**Started**: [Timestamp]
**Agents**: [num_agents]
**Iterations**: 0 / [num_iterations]

---

###### package.json
\`\`\`json
{
  "type": "module",
  "dependencies": {
    "@modelcontextprotocol/sdk": "^1.17.1"
  }
}
\`\`\`

---

## Research State Tracker

###### state-tracker.ts
\`\`\`typescript
// DETERMINISTIC STATE ONLY - No intelligence
export interface ResearchState {
  iteration: number;
  maxIterations: number;
  agents: AgentState[];
  evidenceCollected: number;
  hypothesesGenerated: number;
  hypothesesValidated: number;
  synthesisComplete: boolean;
}

export const state: ResearchState = {
  iteration: 0,
  maxIterations: 3,
  agents: [
    { id: 'agent-1', role: 'literature-scout', status: 'pending', evidenceCount: 0 },
    { id: 'agent-2', role: 'technical-analyst', status: 'pending', evidenceCount: 0 },
    { id: 'agent-3', role: 'synthesis-specialist', status: 'pending', evidenceCount: 0 }
  ],
  evidenceCollected: 0,
  hypothesesGenerated: 0,
  hypothesesValidated: 0,
  synthesisComplete: false
};

// State tracker functions (deterministic only)
export function recordEvidence(agentId: string, count: number): void {
  const agent = state.agents.find(a => a.id === agentId);
  if (agent) {
    agent.evidenceCount += count;
    state.evidenceCollected += count;
  }
}

export function getAgentStatus(agentId: string): string {
  return state.agents.find(a => a.id === agentId)?.status || 'unknown';
}

export function canProceedToIteration(): boolean {
  // Deterministic gate check
  return state.evidenceCollected >= 20 &&
         state.agents.every(a => a.status === 'complete');
}

console.log('State tracker initialized:', state);
\`\`\`

---

## Iteration 0: Initial Exploration

### Agent 1: Literature Scout

###### agent1-iteration0-search.ts
\`\`\`typescript
import { state, recordEvidence } from './state-tracker.js';

// LLM provides search queries (intelligence)
const queries = [
  "[LLM will fill this in]",
  "[LLM will fill this in]"
];

// Agent executes deterministically
const results = {
  phase: 'literature-search',
  agent: 'agent-1',
  iteration: 0,
  queries_executed: queries.length,
  papers_found: 0,  // Will be updated by actual MCP calls
  raw_data: []
};

// State update (deterministic)
recordEvidence('agent-1', results.papers_found);

console.log('Agent 1 status:', results);
\`\`\`

###### agent1-iteration0-findings.ts
\`\`\`typescript
// AGENT 1 FINDINGS (Raw data only, no interpretation)
export const findings = {
  papers: [
    // Will be populated by arxiv-paper-mcp results
  ],
  metadata: {
    count: 0,
    date_range: "",
    categories: []
  }
};
\`\`\`

### Agent 2: Technical Analyst

###### agent2-iteration0-search.ts
\`\`\`typescript
import { state, recordEvidence } from './state-tracker.js';

// LLM provides search queries (intelligence)
const codeQueries = [
  "[LLM will fill this in]"
];

// Agent executes
const results = {
  phase: 'technical-search',
  agent: 'agent-2',
  iteration: 0,
  code_examples_found: 0,
  libraries_found: 0
};

recordEvidence('agent-2', results.code_examples_found);
console.log('Agent 2 status:', results);
\`\`\`

### Agent 3: Synthesis Specialist

###### agent3-iteration0-synthesis.ts
\`\`\`typescript
import { findings as agent1Findings } from './agent1-iteration0-findings.js';
import { findings as agent2Findings } from './agent2-iteration0-findings.js';

// LLM provides synthesis questions (intelligence)
// Agent reads data (deterministic)
const crossAgentPatterns = {
  common_themes: [],  // LLM identifies
  contradictions: [], // LLM identifies
  gaps: []           // LLM identifies
};

console.log('Synthesis complete for iteration 0');
\`\`\`

---

## Iteration 1: Building on Prior Work

[Repeat structure with agents building on iteration 0 findings]

---

## Final Synthesis

###### final-report.ts
\`\`\`typescript
// Aggregates all iterations
import { state } from './state-tracker.js';

const report = {
  research_topic: "[Topic]",
  iterations_completed: state.iteration,
  total_evidence: state.evidenceCollected,
  validated_findings: [],  // LLM provides
  novel_insights: [],      // LLM provides
  recommendations: []       // LLM provides
};

console.log('Research complete:', report);
\`\`\`
\`\`\`

---

## Execution Flow with Sub-Agents

### Iteration 0: Parallel Agent Deployment

**Claude Code spawns 3 sub-agents**:

```bash