At NVIDIA’s GTC 2026 conference, Jensen Huang predicted that every knowledge worker would have 100 AI agents working on their behalf — framing it not as a distant aspiration but as a planning assumption inside NVIDIA itself. The audience applauded. Nobody asked the obvious question: is 100 actually better than 50? Or 20? And at what point does the next agent make everything worse?

The Coordination Tax

In 1967, Fred Brooks documented in The Mythical Man-Month that adding programmers to a late software project makes it later. The reason is mathematical: communication paths between N people grow as N(N-1)/2. A team of 5 has 10 communication paths. A team of 20 has 190. A team of 100 has 4,950.

AI agents don’t have coffee breaks or email inboxes, but they have the same coordination problem. Every agent that reads another agent’s output, shares a knowledge base, or competes for the same context window introduces a communication path. Without structure, 100 agents produce 4,950 potential interference patterns. The result isn’t 100x the intelligence. It’s a cacophony.

This is why most multi-agent demonstrations fail in production. The demo runs 5 agents on a curated task. The production system runs 50 agents on ambiguous real-world data. The demo looks magical. The production system produces contradictory outputs, duplicated work, and hallucinations that no single agent would have generated alone, because the agents are anchoring on each other’s errors without anyone noticing.

A November 2025 McKinsey survey of 1,993 organisations found exactly this pattern: 62% were running AI agent experiments, but only 23% had scaled them at enterprise level. The 39-point gap between experimentation and production scale is almost entirely explained by the coordination problem — not the capability problem.

Two Kinds of Agents

The scaling question becomes tractable when you distinguish between two fundamentally different agent types:

Functional agents do a job — scan news, validate pipeline data, manage calendar conflicts. They operate independently on different tasks. Adding more scales linearly: the 50th research agent covering a new sector doesn’t interfere with the first. The coordination tax is low because they don’t need to talk to each other.

Deliberation agents cross-validate a judgment. One lens assesses technology risk; another assesses market dynamics; a synthesis agent reconciles their disagreements. These agents are interdependent by design — their value comes from interaction. Adding more follows a logarithmic curve, eventually degrading output when the synthesis layer can no longer reconcile competing signals.

Most “100 agents” claims conflate the two. Scaling to 100 functional agents is straightforward — parallel task execution. Scaling to 100 deliberation agents requires coordination architecture that most systems don’t have.

The Hierarchy Solution

The answer, it turns out, is the same one that organisations discovered centuries ago: hierarchy. Not bureaucratic hierarchy that slows things down, but coordination hierarchy that prevents chaos.

Manthan Intelligence runs approximately 43 agents across 8 divisions in four tiers: Directors (8) own functional domains and route tasks; Functional agents (~28) execute jobs within divisions and scale easily; Deliberation agents (~7 active per analysis) cross-validate judgements across 3-9 independent lenses, a range set empirically (3 lenses catch 80% of decision-relevant risks; 9 catch 96%; marginal gain beyond 9 falls under 1%); and a single Synthesis agent compresses all output into one coherent assessment — because synthesis cannot be parallelised.

The Scaling Formula

The practical rule: functional agents scale to demand, deliberation agents scale to diminishing returns, coordination agents scale to hierarchy depth.

A venture fund analysing 30 companies per month needs 6-10 functional agents, 3-9 deliberation agents per analysis, and 3-4 coordination agents — 15-25 active total. An enterprise with 8 divisions needs 3-5 functional agents per division, a deliberation layer for cross-divisional decisions, and 8 Directors plus synthesis — 35-50 active total.

The number 100 isn’t wrong — it’s a milestone, not a destination. An organisation reaches 100 agents when it has enough divisions, enough parallel tasks, and enough coordination infrastructure to keep them productive. Jumping to 100 without the hierarchy is like hiring 100 analysts with no org chart. You don’t get 100x the output. You get chaos with a headcount.

The Question You Should Actually Ask

When someone tells you their system runs N agents, don’t ask “is N big enough?” Ask: what’s the coordination architecture? How do agents avoid duplicating work? How are disagreements resolved? What prevents the 50th agent from degrading the output that 49 agents produced?

If the answer is “they all share a context window” or “they vote” — that’s a system that will break between the demo and production. If the answer describes a hierarchy with clear division of labour, bounded deliberation, and designated synthesis — that’s a system that might actually scale.

The number of agents is a vanity metric. The coordination architecture is the product.

This analysis draws on Fortune’s reporting on Jensen Huang’s NVIDIA GTC 2026 remarks, Fred Brooks’ The Mythical Man-Month, and McKinsey’s State of AI November 2025 report. Human editorial oversight applied.

This analysis is informational and does not constitute investment advice, a research report, or a recommendation to buy, sell, or hold any security.

Charaka Notes by Manthan Intelligence. Subscribe