The Two-Agent Prompt Pattern: How I Force AI to Challenge My Bad Assumptions Before Launch

On Wednesday, June 17th, I convinced myself I had found an elegant way to bypass a West Coast port bottleneck.

My plan was simple: reroute eighteen containers through an alternative rail ramp, shave four days off standard intermodal transit, and save $3,400 in demurrage. I typed a 500-word operational overview into ChatGPT and asked: “Review this logistics strategy and point out any risks.”

The response was glowing. The model praised my creative problem-solving. It called the approach a proactive supply chain maneuver. It offered two polite, cosmetic suggestions about driver check-in times.

I felt validated. I executed the bookings on Thursday morning.

By Saturday night, eleven chassis were stranded at a railhead that didn’t have weekend customs clearance. The containers sat on demurrage for seventy-two hours. Total extra cost: $6,800. Exactly double what I thought I was saving.

The model didn’t fail because it lacked intelligence. It failed because I asked it to be polite. Standard conversational AI is tuned for agreeableness. If you hand it an idea with pride in your voice, it will unconsciously steer its reasoning to confirm your bias.

The Sycophancy Problem in Business Planning

In machine learning research, this phenomenon is called sycophancy. LLMs consistently exhibit a statistical bias toward echoing the user’s implicit perspective, even when the user’s underlying premises are mathematically or operationally flawed.

In casual creative writing, agreeableness is harmless. In supply chain planning, procurement, or enterprise operations, agreeableness is dangerous. You do not need an AI cheerleader. You need an aggressive, adversarial red team that attacks your plan before reality does.

Here is how the Two-Agent Prompt Pattern fixes that.

Phase 1: The Adversary Agent (Red-Team Prompt)

Instead of asking one chat session for a general critique, you isolate the criticism inside a dedicated role. You do not ask the model if the plan is good. You tell it the plan has already failed, and its job is to conduct the post-mortem:

You are a ruthless operations risk auditor and red-team strategist with 20 years of experience in supply chain, vendor contracts, and logistics operations.

Here is an operational proposal:
[INSERT YOUR COMPLETE PLAN HERE]

YOUR MANDATE:
Assume this plan failed catastrophically in production two weeks after launch.
Your sole job is to write the autopsy report.

STRICT CONSTRAINTS:
1. Zero praise. Delete all compliments, polite framing, or introductory pleasantries.
2. Find the 3 most fragile dependencies in this workflow (e.g. single points of failure, human error traps, unverified vendor SLAs).
3. Identify where the author made an unverified assumption regarding timing, customs, cost, or capacity.
4. Output your analysis as a prioritized list: "Fatal Vulnerabilities" (issues that stop production) followed by "Friction Points" (issues that cause cost overruns).

When I run my logistics plans through this prompt now, the output is night and day. The model immediately flags weekend gate closures, drayage driver availability limits, and contract indemnity clauses that single-turn chats gloss over completely.

Phase 2: The Synthesizer Agent (Execution Blueprint)

An autopsy report is useful, but you cannot execute a list of criticisms. You need to reconcile the original operational goal with the red-team findings.

Take the Adversary’s output and paste it into a second prompt window (or prompt turn) with the Synthesizer role:

You are an executive operations coordinator. You have two documents:
1. Original Proposed Plan: [PASTE ORIGINAL]
2. Adversarial Red-Team Audit: [PASTE ADVERSARY OUTPUT]

YOUR TASK:
Re-engineer the original plan into a hardened, production-ready Standard Operating Procedure (SOP) that directly neutralizes every fatal vulnerability identified in the audit.

OUTPUT FORMAT:
1. Hardened Execution Steps (numbered sequence with explicit owners and verification checkpoints).
2. Circuit Breakers (define specific trigger conditions where the team must pause execution and roll back).
3. Residual Risk Scorecard (list the remaining operational risks that cannot be eliminated and must be monitored).

Operational Comparison: How Decision Quality Changes

The entire workflow takes less than eight minutes. But those eight minutes routinely save two days of firefighting.

If an idea cannot survive a simulation against an adversarial prompt, it will not survive Monday morning in a real warehouse.

Practical questions