Are AI Companies Deliberately Teaching Us to Spend More Tokens?

Anthropic published a guide last month, Building agents: When to fold subagents into your main agent loop, and I'll give them credit upfront: it's one of the most honest pieces of agent architecture advice a lab has published.

The argument: most multi-agent setups are Conway's law made executable. Every team ships its own subagent, an orchestrator routes between them, detail dies at every handoff, and one customer measured their single-loop rewrite running about twice as fast as the multi-agent version it replaced.

The fix: fold most subagents back into one main agent loop, load skills and tools on demand ("progressive disclosure"), keep separate agents only for parallel work, fresh-context verification, and modular permissions.

Good advice. Genuinely. If your architecture is a router with six subagents, you should probably read it.
And yet, I finished the article with a question I couldn't shake:

Why does every solution on the menu still involve the model doing the coordination?

Notice what survives the optimization

Strip the article down to its core claim and it's this: the model can now do the orchestration work itself.
Read that sentence again, slowly, and then ask who wrote it. A company that sells tokens. A company whose revenue is, to a first approximation, a function of how many decisions you route through a billed model call.

The multi-agent era taught us to build systems where every routing decision, every handoff, every schema translation was a full model run. That was the sprawl era with a token bonanza, and we all paid for it in invoices and latency.

Now the same industry tells us the cure: don't give the coordination work to many models. Give it to one model, in a loop, forever, with progressive disclosure.

Fewer tokens, sure. But the paradigm is untouched: coordination happens inside a context window, and everything inside a context window is billable.

They sold us the disease. Now they're selling us a managed version of it. The meter was never the issue — the meter is the business model.

Am I claiming a conspiracy? No. I'm claiming misaligned incentives.
Conspiracies require coordination and secrecy. This is simpler and more boring: when your revenue is metered cognition, every architecture question gets answered with "more cognition, please."

Watch what labs benchmark. Reasoning effort. Agentic tool use. Orchestration quality. All of these consume tokens by design.

Now try to find the lab whitepaper titled "Actually, 80% of your agent workloads shouldn't be agentic at all, here's a registry and a routing table."

You won't find it, because the moment coordination stops being an intelligence problem, it stops being a product problem.

Remember:
Fewer tokens, sure. But the paradigm is untouched: coordination happens inside a context window, and everything inside a context window is billable.

Anthropic's own Deep Research feature; which I respect as engineering; burns roughly 15x the tokens of a normal chat interaction, and their own reporting found token spend alone explained most of the performance variance.

That's an honest disclosure. It's also a product strategy: spend more, get better. The slope of that line is the company's income statement.

None of this requires bad faith. It only requires that thousands of smart engineers, optimizing for metrics their employer monetizes, never land on "the cheapest correct answer is the one that never enters the context window."

The number that started this

Here's the comparison that made me write this post.
Take a routine business question, a credit check, a compliance flag, an order status.

In a typical agentic setup, that answer costs you ~27,000 tokens: system prompt, tool schemas, full conversation history resent because the API is stateless, a few rounds of tool calls, thinking tokens padding every turn.

Now ask the same question as what it actually is, a precise, structured query to the system that owns the data. ~700 tokens. Sometimes less.

Same answer. Same correctness. 2.6% of the cost.

Here's the analogy that makes it visceral. Ask someone a simple question in a quiet office: ten seconds of calm attention, clean answer. Now ask the same question in the middle of a packed bar, music blaring, fifty conversations colliding around you. The answer doesn't change. But now you have to shout, repeat yourself, lean in; and the other person burns serious effort just to pick your voice out of the noise. The conversation got harder, not smarter. The effort went to filtering, not to understanding.

A bloated context window is a packed bar. The model isn't thinking harder about your question, it's straining to hear it. And the cruel twist: the audio world spent decades building models to remove that noise; it's literally called the cocktail party problem, and we've quietly rebuilt it, prompt by prompt, at scale. We engineered machines that can isolate a single voice in a stadium, and then we hand them a context window where the stadium never leaves.

The 26,300 tokens you didn't need weren't "intelligence." They were shuffling papers: schemas you weren't calling, history you weren't using, reasoning about routing that a directory could have answered deterministically, the noise floor of the bar. But here's the uncomfortable part, the model bills you the same for thinking and for straining to hear. To the invoice, signal and noise are indistinguishable. That's not a technical limitation. That's a business model.

The question nobody at a lab will ask you

The industry's current debate is: one agent loop, or many subagents?

That's a fake choice. Both answers assume the model should be the coordinator. The real question is:

Should this decision be inside a context window at all?

Discovery, routing, validation, permission checks, we solved this class of problem decades before LLMs existed. Registries. Message passing. Capability-based security. Late binding. None of it requires a 200K-token desk. An agent that asks a lightweight gateway "who can do X?" and receives a route plus a cryptographically signed ticket is doing in zero tokens what a "single agent loop with progressive disclosure" does in a billed reasoning pass.

Full disclosure: this is exactly what I build. I'm the creator of IRC-A, an architecture for decentralized agent networks built around semantic capability routing, prompt rewriting at the delegation boundary, and ephemeral execution tokens.

Flag my conflict of interest accordingly. But the pattern isn't exotic, and it isn't mine: it's what cost-per-task-flat systems look like when you stop using the context window as a coordination mechanism.

A specialist receiving ~200 tokens instead of ~200,000 doesn't just cut the bill — indirect prompt injection dies in the rewrite, because the execution layer never inherits the chat history it would need to be poisoned by.

The point isn't my protocol. The point is that the category, "move coordination out of the intelligence and into the infrastructure", is conspicuously missing from every lab's architecture guidance. And the absence of a category is itself data.

The test I now apply

I'm not claiming the big loop is never right. Genuinely novel reasoning, ambiguous problem decomposition, work where the thinking is the product; pay for the tokens, happily. That's what these models are for.

The test is simpler:

Is the model thinking, or is it shuffling papers? You're billed the same either way. If you can't tell the difference in your own traces, someone else is writing your architecture, and their incentive is the meter.

Twelve years in banking systems teaches you to follow incentives before following advice. The advice we get from labs is often correct, carefully measured, generously shared. And it is never going to end with "the cheapest solution is the one that needs less of what we sell." That's not cynicism. That's just reading the business model.

Progressive disclosure trims the bill. Refusing to coordinate through context abolishes most of it.

Which one do you think gets a migration kit?

I'm Sandro García, a solutions architect with 20+ years in financial systems (Citibank, Bloomberg) and the creator of IRC-A. This is a personal opinion piece; this one is about the economics and the incentives.

Pushback welcome: where does "take coordination out of the model" break down for you? Genuinely asking — the comments are the best part of this platform.