Signal detection is harder than pattern recognition. A pattern is visible in retrospect; a signal is weak, early, and often contradictory with other data you’ve been tracking. The test of a real signal is that three independent sources, using different methodologies, reach the same structural conclusion without coordinating.

Three such signals arrived in the first half of 2026. None are individually new. Together, they point to something that doesn’t yet have a standard name.

Signal One: The Demand Claim

At NVIDIA’s GTC conference in March 2026, Jensen Huang described a future where every knowledge worker would operate with 100 AI agents working on their behalf. The headline circulated and was debated as a vendor boast. The signal is not the number — it’s the context. Huang was not speculating about an abstract future. He was describing NVIDIA’s own workforce planning. A company that builds the infrastructure for AI systems has concluded that the ratio of AI agents to human workers inside its own operations will reach 100:1 within the planning horizon.

Enterprise adoption claims from vendors are routinely overstated. What is less common is a company of NVIDIA’s scale making 100:1 a planning assumption rather than a marketing aspiration. That moves the signal from promotional to structural.

Signal Two: The Practitioner Inflection

In May 2026, Andrej Karpathy published a summary of his Sequoia AI Ascent fireside chat that included a deceptively simple observation: in December 2025, over roughly one month, his personal workflow shifted from writing 80% of his own code to having 80% of it generated by AI systems. This was not a planned transition. It happened incrementally and then suddenly — the classic inflection point structure.

Karpathy’s observation matters not because it applies universally, but because of who he is. He is the practitioner most closely associated with drawing the actual line between human-irreplaceable and AI-automatable tasks in software development. When the person who has most credibly articulated that boundary reports that his own practice has crossed it, that is not a marketing claim. It is a field report from the frontier.

The inflection Karpathy described is not a productivity improvement. It is a qualitative shift in the division of cognitive labour. That shift, if it generalises beyond software, implies a different infrastructure requirement than current enterprise tooling assumes — you need a coordination layer that manages agents on your behalf, not just a tool that you operate.

Signal Three: The Deployment Gap

McKinsey’s State of AI survey in November 2025, covering 1,993 organisations, found that 62% of enterprises were running AI agent experiments while only 23% had scaled them at the enterprise level. The 39-point gap between experimentation and production scale is not primarily a capability problem — the models that work in demos work in production. It is a coordination problem. The tooling for deploying AI agents at scale, managing interactions between them, allocating tasks coherently, and maintaining reliable outputs across systems does not yet exist in a form enterprises can adopt without significant custom engineering.

A 39-point chasm between experiment and scale is the signature of a market waiting for infrastructure. It is the same gap that existed in the early enterprise cloud market before orchestration tooling became standard: the technology worked, but deployment at scale required a layer that hadn’t been commoditised yet. The gap closed when the orchestration layer arrived. The 62%/23% split in AI agent adoption will close the same way.

What the Three Signals Converge On

The three signals are methodologically independent: one comes from a hardware infrastructure company’s internal planning, one from a practitioner’s personal field observation, and one from a large-scale enterprise survey. They reach the same structural conclusion: the bottleneck in AI agent deployment is not model capability — it is the coordination layer.

This layer — whatever it ends up being called — needs to handle task routing, output validation, context sharing, budget management, conflict resolution, and synthesis. It does not yet exist as a standard product. Every organisation that has achieved production-scale agent deployment has built something like it themselves. That is the characteristic of a market one to two years away from infrastructure commoditisation.

The Charaka View

Manthan’s 46-agent operating infrastructure is an early instance of this coordination layer. The architecture — division directors for routing, functional agents for parallel execution, deliberation agents for cross-validation, synthesis agents for compression — maps directly onto the problems the three signals identify. We built it because we needed it; off-the-shelf orchestration tools did not resolve multi-agent coordination at the depth our pipeline requires. Every organisation deploying agents at scale will need some version of this. The ones building it now, rather than waiting for a commodity solution, will have learned something the vendor version cannot teach.

This analysis draws on Fortune’s reporting on Jensen Huang’s NVIDIA GTC 2026 remarks, Andrej Karpathy’s Sequoia AI Ascent 2026 summary, 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