The Language Trap

Why AI agents are exhausting themselves speaking English — and what it means for you.

Part 1 – The explosion nobody saw coming

Have you seen what's happening right now?

The number of AI agents on the internet is growing at a rate I wouldn't have imagined six months ago. We're moving from a handful of experiments to a full-blown army of tiny digital entities capable of acting on our behalf.

The promise is dizzying.

These agents could book the cheapest plane ticket, manage your schedule 24/7, submit insurance claims, or continuously scan a codebase for vulnerabilities and patch them in real time.

It sounds beautiful.

Except reality is rougher. The technology is still young. It's improving fast, that's true, but it remains crude.

And alongside these promises, we see headlines every day: spam, security breaches, cascading system failures.

Part 2 – The vacation nightmare

The problem gets worse when you move from a single agent to a team.

Take a simple example. Imagine you delegate your holiday planning to two AI agents.

The first handles the flight. It "hallucinates" – that's the technical term – a cheaper airport. Except that airport is 400 miles away from your actual destination.

The second handles the hotel. It offers you a "super cheap" room nearby. Except in the language of agents, "super cheap" often means "non-refundable."

Result: you have a non-refundable room. That you will never see.

This isn't a joke. It's a concrete case that illustrates a fundamental problem: agent coordination is a nightmare.

Part 3 – The standard method that doesn't work

Faced with this observation, researchers proposed an approach that, at first glance, looks like what we all do.

One agent writes a plan. A second critiques it. A third solves the problem.

On paper, it's clean. It's structured. It's what everyone does with agents today.

But there's a detail that caught my attention.

Most agents communicate with each other just like us: with words.

Sentences. Tokens. Natural language.

And I had this intuition: why?

Why would artificial entities, which aren't constrained by a mouth or vocal cords, insist on speaking English?

Part 4 – The alphabet wasn't built for thinking

I remembered a demonstration I had seen somewhere. A neural interface that turns thoughts into text. You think of a letter of the alphabet, and it magically appears on screen.

It's fascinating.

But upon reflection, a question arises: why use the alphabet?

The alphabet is optimized for writing. For transcription. For written communication between humans.

But is it optimized for thinking?

No.

And what if we applied this reasoning to AI agents?

Part 5 – The invisible bottleneck

I observed how a classic multi-agent system works.

The first agent does its work. It packages the result. It passes it to the next one. The second does the same. The third too.

So far, nothing unusual.

But look closely at what happens during each transmission.

The agent must write complete sentences. It decodes tokens one by one. It structures its thoughts in a linear, grammatical format. The next agent must read these sentences, interpret them, decode them, then re-encode all the information into its own internal representations.

It's a massive bottleneck.

A considerable amount of time and energy is wasted in this permanent translation.

I asked myself: who decided that agents should speak English?

Part 6 – The idea that changes everything

And then I discovered an approach that knocked me off my chair.

Forget English. Forget letters entirely. Let's link up their brains.

Not literally, obviously. But the idea is radical:

Instead of exchanging English words, the agents pass raw numbers to each other. Undecoded signals. What's called cross-agent latent state transfer.

In practice, they send their internal states directly to one another. No translation. No formatting. No loss.

The results are staggering.

Three agents communicating this way can work together over multiple rounds. They refine their answers progressively. And they do it much cheaper than agents that insist on speaking English.

With the same amount of computation, you get better answers.

Part 7 – The numbers that speak for themselves

I'll give you the figures as I saw them.

First result:

On competition-level math problems, the success rate jumps from 73% to 86%.

Thirteen points gained.

And this isn't with massive, expensive models. These are sub-10 billion parameter models. Free, accessible models.

Second result:

Token usage drops by 75%.

The agents essentially "evaporated" into latent space. They exchange the essential information without the overhead of natural language.

What this means: smaller models, consuming far fewer tokens – and therefore far less money – achieve performance that puts them within striking distance of much larger, much more expensive systems.

Part 8 – The cost of it all

When I saw these numbers, I thought: "This must cost a fortune to train."

Four dollars.

Yes. Four dollars.

That's the price of a coffee.

With this ridiculously small budget, these agents perform at a level that, just a year ago, was reserved for the most expensive systems.

Part 9 – A new dynamic

This approach might even reveal a new scaling law.

More rounds of latent communication, better results.

Performance continues to improve as you increase the number of raw exchanges between agents. This isn't just a technical optimization; it's a paradigm shift.

Part 10 – The good teacher trap

I asked myself a subtle question.

The training for each agent is provided by a giant AI model. So if the agents perform well, one might wonder:

Is it better because of latent communication, or is it simply because the teacher is excellent?

In other words: does the performance come from the architecture or from the content being taught?

The researchers had the same intuition. So they ran a controlled comparison:

They gave the same teacher to other architectures. And the new architecture still outperformed.

Conclusion: brain-linking really works. The performance comes from how agents communicate, not just from what they were taught.

Part 11 – The limits (I'm being honest)

I'm not going to sell you a dream without nuance. There are limits.

First limit: scale.

The tests were run on smaller models. We don't yet know if these results hold for larger ones.

If it doesn't scale, we still have small models on steroids. That's very useful for embedded or low-cost applications.

If it does scale, it's a complete upheaval of the landscape.

Second limit: thought length.

There's an optimal length for "latent thought": about 80 steps. Beyond that, returns diminish.

It's still respectable – these agents already solve Olympiad-level math problems. But it's a constraint to keep in mind.

Part 12 – What this means for you

If you work in AI, the lesson is clear.

Optimization doesn't always come through model size.

You have an alternative to the frantic race for more parameters. By rethinking how your agents exchange information, you can boost performance and drastically cut computational costs.

The question is no longer just "which model should I use?" but "how do they exchange information?"

If you're a marketer, the analogy is direct.

You handle raw data streams: customer feedback, interview transcripts, support logs, reviews, comments.

You tend to "clean" them. Format them. Turn them into presentable sentences.

That's a mistake.

The hesitations, the repetitions, the awkward phrasing – that's your "latent space." That's where the real intentions, the real objections, the real purchase triggers are hiding.

When you translate all of this into clean marketing language, you lose the signal. You're making the same mistake as agents exhausting themselves encoding and decoding English tokens.

Keep it raw.

Part 13 – The current state

Let me be clear: this isn't production-ready yet.

The code and models are available. But it's still rough, still early. It shows immense potential, but you can't just plug this in and expect everything to work immediately.

It's research. Promising research, but research nonetheless.

Conclusion

What makes this approach so exciting isn't just the performance gain.

It's that it challenges an assumption we hadn't even identified as one.

The idea that natural language is the best communication format for agents.

It's wrong.

Formatted, clean, written language is a tool. Useful, certainly. But sometimes, the most valuable thing is what hasn't been translated yet.

The raw. The undecoded. The latent space.

Keep that in mind.