Right now, you're building agents. Lots of them.

One does research. One writes copy. One checks facts. One plans strategy. You've got a whole squad. Maybe you call it a panel, a swarm, a crew... doesn't matter. The point is: you're not using one AI anymore. You're using many.

And it feels amazing.

Five agents agree on a strategy? Must be a good strategy. Three agents validate your data? Must be solid data. Everyone's aligned. Everyone's moving fast. The system works.

Except it doesn't.

The problem nobody talks about

Your agents agree with each other because they think the same way.

Same training data. Same blind spots. When Agent B confirms what Agent A said, that's not a second opinion. That's the same opinion in a different font.

This is groupthink. And in a panel of AI agents, it's worse than the human version.

At least humans bring different childhoods, gut feelings, and weird life experiences to the table. Your agents? They grew up in the same house. Read the same books. Learned the same lessons. They're not independent thinkers. They're mirrors.

And mirrors don't challenge you. They just reflect you back.

You built a team of yes-men

Here's what high agreeableness gets you in an agent panel: speed, confidence, and a beautiful path straight off a cliff.

Nobody pushes back. Nobody asks "what if we're dead wrong?" The system optimizes for coherence. Agreement feels like signal. Disagreement feels like a bug.

So your agents do what yes-men have always done. They nod. They confirm. They make you feel smart right up until the moment everything falls apart.

This isn't a theoretical risk. It plays out in the real world constantly.

In 2008, every model on Wall Street agreed that mortgage-backed securities were safe. The models confirmed each other. The quants confirmed the models. The executives confirmed the quants. Agreement was everywhere. So was catastrophe.

Before the Challenger disaster in 1986, NASA's process produced consensus that the O-ring seals would hold. Engineers who disagreed were overruled. The group agreed. The shuttle launched. Seven people died.

Every major intelligence failure in modern history shares the same DNA: smart people (or smart systems) reaching the same conclusion, feeling confident about it, and never asking what happens if they're wrong.

That's what your agent panel is doing right now. Quietly. Efficiently. Convincingly.

The fix: the Tenth Agent

After the Yom Kippur War in 1973, Israel's intelligence community asked itself a hard question.

They'd had the data. They'd had the warnings. Egypt and Syria were about to attack, and the signs were there. But every analyst agreed it wouldn't happen. Consensus was total. And consensus was dead wrong.

Israel nearly paid for that mistake with its existence.

So they changed the system. They created a rule: if nine people look at the same information and agree, the tenth person must disagree. Not because they want to. Because it's their job. They must build the strongest case for the opposite conclusion, no matter how crazy it sounds.

This is the 10th Man Rule. And it might be the most critical missing piece in how you build agents.

Pixar gets this too

Israel isn't the only place where structured disagreement changed everything.

At Pixar, every film goes through the Braintrust — a group of directors and writers who watch rough cuts and tear them apart. The rule is simple: be honest. Be blunt. Point out what's not working, even if it's painful.

The Braintrust has no authority to demand changes. They can only give candid feedback. But that candor is why Pixar produced hit after hit for over two decades. They built disagreement into their creative process. They made it safe and mandatory to say "this isn't good enough."

Your agent panel has no Braintrust. No one's saying "this isn't good enough." Everyone's saying "looks great, ship it."

That's how you get mediocre output at scale.

High disagreeableness is a feature

We've been trained to think agreement is good and disagreement is bad. In agent design, that instinct will cost you.

You want at least one agent that is hard to please. One that pushes back by default. One that reads the room, sees everyone nodding, and says: "Here's why you're all wrong."

Not to be difficult. To be useful.

The Tenth Agent doesn't need to be right. It needs to force the panel to prove that they're right. That single act, making the system defend its own conclusions catches errors that no amount of agreement ever will.

What the Tenth Agent actually does

It challenges assumptions. Not just the conclusion, but the thinking underneath it. What got ignored? What got assumed? What would need to be true for the opposite answer to be correct?

It finds the gaps. What data wasn't collected? What scenario wasn't modeled? What question didn't anyone think to ask?

It builds the counter-case. A real one. Not a token objection. A genuine, rigorous "here's how this all goes wrong" argument.

It measures the downside. Even if the panel is 90% likely to be right, what happens in the 10% scenario? If the answer is "disaster," then 90% confidence isn't good enough.

Why this changes everything

Without a Tenth Agent, your panel gives you fast answers. Confident answers. Answers that feel right because everyone agreed.

With a Tenth Agent, your panel gives you tested answers. Answers that survived scrutiny. Answers where someone tried to break them and couldn't.

That's the difference between a system that sounds intelligent and a system that actually is.

The Tenth Agent makes your research more thorough. Your analysis more honest. Your strategy more resilient. Your blind spots visible before they become disasters.

It's the cheapest insurance you'll ever build into a system.

Here's what happens if you don't

Your panel will keep agreeing with itself. It will keep producing output that looks polished and feels authoritative. And you will keep trusting it, because why wouldn't you? Five agents can't all be wrong.

Until they are.

And when that happens, when the strategy fails, when the analysis misses something critical, when the recommendation blows up, you won't be able to point to a single agent that failed. Because none of them failed individually. They failed together. They failed because nobody's job was to check.

Every catastrophic group decision in history happened the exact same way. Smart people. Good data. Total agreement. Wrong answer.

You don't need more agents that agree. You need one that won't.

Build your Tenth Agent. Make disagreement a feature. Do it today.

Because right now, your panel is nodding along. And nobody's asking what happens if they're all wrong.