The Multi-Agent Trap Nobody Talks About

A Stanford research team fed the same clinical case to five GPT-4 instances. Then let them discuss.

By round three, all five had converged on the same diagnosis.

It was wrong.

The first agent’s confident-but-incorrect assessment anchored the others. Every round of “deliberation” destroyed analytical diversity instead of building it. They didn’t have a deliberation chamber — they had an expensive consensus machine.

This is the central failure mode of multi-agent AI. And it’s almost never discussed.

Language models are trained to produce coherent, agreeable text. Put three agents in a discussion thread and ask them to reach a conclusion — they will. It won’t be the most accurate conclusion. It’ll be the most agreed-upon one. That’s a fundamentally different thing.

When we designed Manthan Intelligence’s Analytical Council, the first architectural rule was: agents cannot see each other’s work before committing their own conclusions.

The architecture runs in three phases:

→ Phase 1: Independent assessment. Each analytical lens evaluates in complete isolation. No agent sees what any other agent concluded. Same source data. Zero cross-contamination.

→ Phase 2: Lock, then reveal. Every lens commits its verdict before synthesis begins. These commitments are immutable — no lens revises after seeing what the others said.

→ Phase 3: Synthesis under tension. The synthesis layer looks for productive disagreement — not consensus. When two lenses reach opposite conclusions from the same data, that tension is the most valuable output in the analysis. Not something to smooth over.

The Knowledge Graph is the only shared resource. It contains facts — funding rounds, market sizes, benchmarks. Zero opinions. Zero prior verdicts. Shared context, independent judgment. Deliberately.

Multi-agent value comes from genuine analytical diversity. The moment agents anchor on each other, you have an expensive agreement machine.

The hardest part of building a multi-agent system isn’t making agents talk to each other. It’s preventing them from talking too early.

The full architecture behind this — including what the Knowledge Graph does, and what it deliberately doesn’t do — is in Charaka Notes #027.

Read more at getmanthan.com