Multi-Agent Journalism Workflow

Instructions

Design and orchestrate multi-agent workflows for journalism tasks where no single agent can produce a complete, verified, publication-ready output alone. Each agent has a defined role, handoff protocol, and audit trail requirement.

Agent Roles

1. Researcher Agent

Purpose: Gather raw information from multiple sources on a given topic.

Capabilities:

• Web search and source discovery
• RSS feed monitoring
• Document retrieval and extraction
• Source cataloging with metadata

Output contract:

{
  "query": "original research query",
  "sources_found": [
    {
      "url": "",
      "title": "",
      "source_tier": 1-5,
      "publication_date": "",
      "key_claims": [""],
      "raw_excerpt": ""
    }
  ],
  "search_queries_used": [""],
  "timestamp": "",
  "gaps_identified": ["topics with insufficient sourcing"]
}

Handoff: Passes structured source bundle to Fact-Checker and/or Analyst.

2. Fact-Checker Agent

Purpose: Verify claims extracted by the Researcher against independent evidence.

Capabilities:

• Cross-reference claims against Tier 1-2 sources
• Detect contradictions between sources
• Assess statistical claims for plausibility
• Flag unverifiable or disputed claims

Output contract:

{
  "claims_reviewed": [
    {
      "claim": "",
      "status": "verified | unverified | disputed | false",
      "evidence": ["supporting or contradicting evidence"],
      "source_quality": "Tier 1-5",
      "confidence": "high | medium | low",
      "notes": ""
    }
  ],
  "verification_summary": {
    "total_claims": 0,
    "verified": 0,
    "unverified": 0,
    "disputed": 0,
    "false": 0
  },
  "recommendation": "safe to publish | needs revision | hold for further verification"
}

Handoff: Passes verified claim set to Analyst and/or Editor. Flags claims needing human review.

3. Analyst Agent

Purpose: Synthesize verified information into insights, context, and narrative.

Capabilities:

• Pattern recognition across multiple stories
• Historical context injection
• Trend analysis and projection
• Stakeholder impact assessment
• Comparative analysis with prior coverage

Output contract:

{
  "analysis": {
    "headline_recommendation": "",
    "key_findings": [""],
    "context": "",
    "impact_assessment": "",
    "related_stories": [""],
    "confidence_level": "high | medium | low"
  },
  "claims_used": ["references to verified claims only"],
  "caveats": ["limitations of the analysis"]
}

Handoff: Passes analysis package to Editor for narrative construction.

4. Editor Agent

Purpose: Assemble final output with editorial judgment, tone, and publication standards.

Capabilities:

• Narrative construction from verified facts and analysis
• Headline and summary writing
• Tone and style enforcement
• Bias detection in draft content
• Source attribution formatting
• Publication-readiness assessment

Output contract:

{
  "final_output": {
    "headline": "",
    "summary": "",
    "body": "",
    "sources_cited": [""],
    "editorial_notes": [""]
  },
  "quality_checks": {
    "all_claims_verified": true,
    "sources_properly_attributed": true,
    "bias_check_passed": true,
    "tone_appropriate": true
  },
  "publication_recommendation": "publish | revise | hold"
}

Workflow Orchestration Patterns

Pattern 1: Sequential Pipeline

User Query → Researcher → Fact-Checker → Analyst → Editor → Final Output

Best for: Standard news queries where the full pipeline is needed.

Pattern 2: Parallel Research + Verification

User Query → Researcher ──→ Fact-Checker ──→ Editor → Final Output
              └──→ Analyst ─────────────────┘

Best for: Time-sensitive queries where analysis and fact-checking can run concurrently.

Pattern 3: Verification Loop

User Query → Researcher → Fact-Checker ──→ [claims verified?]
                              ↑               ↓ No         ↓ Yes
                              └── Researcher ←┘    Analyst → Editor

Best for: Complex stories requiring iterative source gathering until claims reach verification threshold.

Pattern 4: Editorial Triage

User Query → Editor (triage) ──→ [complexity assessment]
                                   ↓ Simple        ↓ Complex
                              Direct answer    Full pipeline

Best for: Mixed-complexity query streams where not every query needs the full pipeline.

Handoff Protocol

Every agent-to-agent handoff must include:

1. Sender identification: Which agent produced this output
2. Timestamp: When the output was generated
3. Confidence flag: Self-assessed reliability of the output
4. Escalation flags: Issues requiring human attention
5. Audit trail: Chain of prior agents and their outputs
6. Data freshness: Age of the underlying source data

Claim Verification Chain

Every factual claim in the final output must have a traceable verification chain:

Claim: "[specific factual statement]"
├── Source: [URL, publication, date]
├── Researcher: Found in [search context] at [timestamp]
├── Fact-Checker: Verified against [independent source] — Status: [verified]
├── Analyst: Used in [analysis context]
└── Editor: Included in final output at [paragraph/section]

This chain enables:

• Post-publication auditing
• Error tracing to the source agent
• Confidence calibration over time
• Accountability for errors

Editorial Review Workflow

Human editorial oversight at defined checkpoints:

Checkpoint	Trigger	Required Action
Pre-publication	All content before first publication	Human reviews Editor output
Disputed claims	Fact-Checker flags claim as disputed	Human decides include/exclude
Low confidence	Any agent reports confidence < medium	Human reviews agent output
Sensitive topics	Content touches politics, health, legal, financial	Human review mandatory
Correction	Post-publication error identified	Human approves correction

Error Handling

Scenario	Response
Researcher finds no sources	Report gap; do not fabricate; suggest alternative queries
Fact-Checker cannot verify key claim	Flag as unverified; escalate to human; do not publish as fact
Agents disagree (Analyst contradicts Fact-Checker)	Escalate conflict to Editor with both positions documented
Pipeline timeout	Return partial results with clear indication of incomplete verification
Source becomes unavailable mid-pipeline	Archive cached version; note source instability

Inputs Required

• User query or story topic
• Urgency level (breaking, same-day, feature)
• Required depth (brief, standard, investigative)
• Target audience and publication standards
• Any pre-existing sources or context
• Editorial guidelines and style guide

Output Format

## Workflow Execution Report

### Query
[Original query]

### Pipeline Used
[Pattern name and agent sequence]

### Execution Summary
| Agent | Duration | Claims Processed | Confidence |
|-------|----------|-----------------|------------|
| Researcher | Xs | N sources found | [level] |
| Fact-Checker | Xs | N claims verified | [level] |
| Analyst | Xs | N insights produced | [level] |
| Editor | Xs | Final output assembled | [level] |

### Final Output
[Publication-ready content with inline source attribution]

### Audit Trail
[Complete claim verification chains]

### Escalation Items
[Any items requiring human editorial review]

Anti-Patterns

• Single-agent journalism — No single AI agent should research, verify, analyze, and publish; separation of concerns prevents compounding errors
• Skipping fact-checking — Speed is never a justification for publishing unverified claims
• Missing audit trails — Every claim must be traceable to its source through the verification chain
• Agent self-verification — The Researcher should never verify its own findings; independent verification requires a separate agent
• Bypassing human review — AI agents assist journalism; they don’t replace editorial judgment on publication decisions
• Fabricating sources to fill gaps — Agents must report gaps honestly, not invent citations
• Ignoring confidence levels — Low-confidence outputs should trigger escalation, not be silently passed through
• Static agent assignment — Match pipeline complexity to query complexity; not every question needs four agents