The "Double-Agent" Workflow: How to Eliminate 95% of AI Errors by Separating the Maker from the Verifier

Using lightweight sub-agents to catch hallucinations before they hit your inbox.

When I first started consulting in the AI space, I fell into the same trap everyone else does: I thought the solution to a complex business problem was just a “longer, better prompt.” I’d spend three hours writing a 2,000-word master prompt, hit run, watch the AI spit out a massive wall of text, and pray to the tech gods that it didn’t hallucinate some crucial detail.

It was exhausting. It was fragile. And frankly, it was a terrible way to run a business.

A few months ago, I was brought in to help Reagan who runs a consulting firm. She helps mid-sized financial services firms navigate the absolute minefield of state-by-state regulatory filings.

It’s high-stakes, low-margin-for-error work. If Reagan’s team misses a single state-specific exemption rule or filing deadline, her clients face five-figure fines, and her reputation goes up in smoke.

Reagan, being an enterprising business owner, tried to automate her initial regulatory audit reports using a single, massive Claude prompt. She fed the AI the client’s financial profile and the state regulations, and asked it to draft the audit.

For three weeks, she thought she was a genius.

Then came the 2 AM phone call.

---

Build and launch custom, white-labeled web applications from your AI workflows in under 10 minutes with $0 hosting bills and zero code.

---

The AI had hallucinated a critical SEC filing exemption for a major client. Because the model got “lazy” halfway through its own massive output, it simply assumed a rule applied when it didn’t. Reagan missed it because she was skim-reading the 15-page report the AI generated.

They caught the error with about four hours to spare before a $15,000 regulatory fine kicked in. Reagan was ready to go back to doing everything manually on spreadsheets.

The mistake Reagan was making was trusting AI to grade its own homework.

The “Double-Agent” Workflow

In cognitive psychology, and now in advanced AI engineering, we talk about self-preferential bias. If you ask an AI model to generate a complex document, and then in that same chat window you ask, “Hey, can you double-check this for errors?” the model will almost always say, “Yep, looks perfect!”

It is blind to its own reasoning trail. It wants to be consistent with what it already wrote.

The breakthrough is simple: You must separate the Maker from the Verifier.

Instead of one giant prompt trying to do everything, we split the job between two distinct, lightweight roles.

1. The Maker: A model optimized for drafting, writing, and synthesizing raw data.

2. The Verifier: A completely separate, fresh model instance (or a different model entirely) that has no idea how the Maker came to its conclusions. It only sees the final draft and a strict compliance rubric.

[Raw Client Data] 
       │
       ▼
┌──────────────┐
│  THE MAKER   │ ◄────────────────────────┐
└──────┬───────┘                          │
       │ (Draft Report)                   │ (Structured Corrections)
       ▼                                  │
┌──────────────┐                          │
│ THE VERIFIER │ ──(Fails Rubric?)────────┘
└──────┬───────┘
       │ (Passes Rubric)
       ▼
[Final Approved Report]

Now, I’ll be honest with you: if you’re just doing this manually in ChatGPT or Claude, you can just open two separate tabs. But if you actually try to automate this behind the scenes using Make.com or Zapier, you’re going to run into a massive, frustrating roadblock that I had to learn the hard way.

The first time I set up a multi-agent loop for Reagan, I woke up the next morning to a $120 API bill.

Why? Because the Maker and the Verifier got into a passive-aggressive, infinite loop. The Maker would write something, the Verifier would reject it, the Maker would “fix” it but break something else, and the Verifier would reject it again. They did this back and forth, hundreds of times, while I was fast asleep.

To build a system that actually works on autopilot, we need two things: structured JSON outputs so our automation tools can read the results, and a “three-strike” circuit breaker to prevent infinite loops.

---

The Copy-Paste Assets