Your agent needs a fact-checking buddy (or it'll confidently lie to you forever)

Your agent just told you the quarterly revenue numbers are up 15%. You made the slide deck. Sent it to your boss. Then you found out it was completely wrong.

This is the agent hallucination problem nobody talks about: it's not the obvious nonsense that kills you. It's the plausible-sounding facts that slip through because they feel right.

Here's the pattern that catches these before they escape: give your agent a fact-checking buddy.

The Two-Agent Verification Pattern: Agent A does the work. Agent B verifies the facts. Agent A only ships after Agent B signs off.

I learned this the hard way when my business agent told a client we had 47 case studies. We had 12. The client asked to see all 47. Awkward.

Here's how I built the verification loop:

# Primary agent completes task
result = primary_agent.execute(task)

# Verification agent checks factual claims
facts_to_verify = verifier_agent.extract_claims(result)
verification_report = verifier_agent.verify_facts(facts_to_verify)

# Only ship if verification passes
if verification_report.confidence > 0.85:
    return result
else:
    return primary_agent.revise(result, verification_report.issues)

The verification agent has three jobs:

• Extract factual claims — numbers, dates, names, statistics
• Cross-reference against known sources — your CRM, financial data, previous reports
• Flag confidence levels — "I can verify this" vs "I can't find supporting data"

The magic is in the confidence threshold. Set it too high and nothing ships. Too low and bad facts slip through. I found 0.85 works for most business content.

Here's the verification prompt I use:

Extract all factual claims from this content:
- Numbers and statistics
- Dates and timelines  
- Company/product names
- Process descriptions

For each claim, respond with:
- VERIFIED: I found supporting evidence
- UNVERIFIED: No supporting evidence found
- CONTRADICTED: Evidence suggests this is wrong

Only mark VERIFIED if you have direct evidence.

The verification agent doesn't need to be smart — it needs to be skeptical. I use a cheaper model (GPT-4o mini) with instructions to be paranoid about facts.

This catches three types of hallucinations:

• Fabricated numbers — "We processed 1,247 orders" when you processed 124
• Wrong dates — "Last quarter" when referring to data from six months ago
• Invented features — "Our API supports webhooks" when it doesn't

Warning: This pattern doubles your API costs for verified content. But one wrong fact in a client presentation costs way more than a few extra API calls.

The verification agent also learns. When it flags something as unverified and you confirm it's actually correct, that goes into its memory for next time.

I run this pattern for anything that leaves my company — client reports, marketing copy, financial summaries. Internal notes and drafts skip verification to save costs.

The result? My agents still hallucinate, but the hallucinations never make it to production. The verification buddy catches them at the gate.