CURRENT | Market Pulse

The Stratification Problem

Every Layer Works. Nobody Owns the Outcome.

By Rina Takahashi— March 19, 2026

Feature image for article: Every Layer Works. Nobody Owns the Outcome.

At GTC last week, NVIDIA unveiled kernel-level sandboxing for agent runtimes. In the same window, Google and Microsoft pushed WebMCP into W3C incubation. Visa and Mastercard each advanced competing protocols for agent-initiated payments. Three layers of the agent stack, three governance models, each technically serious. Each also scoped to exactly what its owner can control. When an agent researches, interacts, and transacts across all three layers, who's accountable for the outcome? The dashboards all look green.

The Stratification Problem

Every Layer Works. Nobody Owns the Outcome.

By Rina Takahashi— March 19, 2026

At GTC last week, NVIDIA unveiled kernel-level sandboxing for agent runtimes. In the same window, Google and Microsoft pushed WebMCP into W3C incubation. Visa and Mastercard each advanced competing protocols for agent-initiated payments. Three layers of the agent stack, three governance models, each technically serious. Each also scoped to exactly what its owner can control. When an agent researches, interacts, and transacts across all three layers, who's accountable for the outcome? The dashboards all look green.

This Week's Stack

A lot of plumbing went in this week. Runtime sandboxes from NVIDIA, browser-level protocol support for WebMCP, competing payment rails from Visa and Mastercard, creative orchestration from Luma. The infrastructure layers are arriving fast, often well ahead of the governance questions they'll eventually surface.

On the research side, a useful corrective: those million-token context windows degrade badly long before they hit their advertised ceilings. Worth keeping in mind when someone pitches you an agent that reads your entire codebase.

Here's what landed, at a glance.

This Week's Stack

A lot of plumbing went in this week. Runtime sandboxes from NVIDIA, browser-level protocol support for WebMCP, competing payment rails from Visa and Mastercard, creative orchestration from Luma. The infrastructure layers are arriving fast, often well ahead of the governance questions they'll eventually surface.

On the research side, a useful corrective: those million-token context windows degrade badly long before they hit their advertised ceilings. Worth keeping in mind when someone pitches you an agent that reads your entire codebase.

Here's what landed, at a glance.

Agent Infrastructure

NVIDIA Unveils Full Agent Runtime Stack at GTC

GTC brought the OpenShell runtime, CPU-optimized Vera Rubin racks for agent workloads, and an Agent Toolkit with 15+ enterprise partners including SAP, Salesforce, and ServiceNow. The NemoClaw/OpenClaw security framing positions NVIDIA as the sandbox layer for enterprise agents, though NemoClaw remains alpha-stage.

Web Protocols

WebMCP Now Testable in Chrome Canary Build

The WebMCP protocol, co-developed by Google and Microsoft, is live behind a flag in Chrome Canary. It lets agents interact with websites through declared structured APIs instead of pixel-scraping. The spec is incubating through a W3C community group, with formal browser announcements expected mid-to-late 2026.

Agentic Payments

Visa and Mastercard Ship Competing Agent Protocols

Both networks launched agentic commerce frameworks in the same quarter, each building on Cloudflare's Web Bot Auth cryptographic layer. Visa is coordinating with 100+ partners and aligning with OpenAI and Coinbase standards. Mastercard's Agent Suite arrives Q2. Liability allocation remains conspicuously unaddressed.

Model Limits

Context Rot Research Punctures Long-Context Window Claims

Research from Chroma and others finds models advertising 1M+ token context windows suffer severe performance degradation well before their limits, with effective capacity typically topping out around 60–70% of the headline number. Semantically similar distractors make things worse, which is bad news for agentic coding tasks.

Frontier Models

Claude Opus 4.6 Ships With Million-Token Window

Anthropic's latest Opus release brings a generally available 1M token context window and leads Terminal-Bench 2.0 for agentic coding at 65.4%. A new context compaction feature automatically summarizes older context to sustain longer sessions. Parallel subagent teams arrive as a research preview.

Creative Agents

Luma Agents Orchestrate Multi-Model Creative Production Pipelines

Luma launched Agents, a platform coordinating 8+ models across text, image, video, and audio from a single brief. Publicis Groupe, Adidas, and Mazda are already running on it. The CEO's claim that one campaign dropped from $15M to under $20K remains unaudited.

The Transaction Layer

Visa and Mastercard Ship Competing Agent Payment Protocols While the Liability Framework Remains Unwritten

Both card networks shipped agentic commerce frameworks in the same quarter. Visa's Intelligent Commerce spans over 100 partners. Mastercard's Agent Pay completed Australia's first authenticated agent transaction in January. The protocol infrastructure is moving.

The liability infrastructure is standing still. Visa's Ramachandran has said openly that agent-initiated disputes remain unresolved. Traditional chargebacks involved four parties. Agents introduce a fifth, and nobody has published rules for how responsibility flows across that longer chain.

Then Citrini Research published a thought experiment about agents routing around interchange fees, and billions briefly evaporated from both networks' market caps. The scenario was speculative. The sell-off was real. What rattled investors wasn't the specific prediction so much as the sudden visibility of a gap: the payments layer is building fast on top of questions it cannot yet answer about who pays when an agent gets it wrong.

The Transaction Layer

Visa and Mastercard Ship Competing Agent Payment Protocols While the Liability Framework Remains Unwritten

Both card networks shipped agentic commerce frameworks in the same quarter. Visa's Intelligent Commerce spans over 100 partners. Mastercard's Agent Pay completed Australia's first authenticated agent transaction in January. The protocol infrastructure is moving.

The liability infrastructure is standing still. Visa's Ramachandran has said openly that agent-initiated disputes remain unresolved. Traditional chargebacks involved four parties. Agents introduce a fifth, and nobody has published rules for how responsibility flows across that longer chain.

Then Citrini Research published a thought experiment about agents routing around interchange fees, and billions briefly evaporated from both networks' market caps. The scenario was speculative. The sell-off was real. What rattled investors wasn't the specific prediction so much as the sudden visibility of a gap: the payments layer is building fast on top of questions it cannot yet answer about who pays when an agent gets it wrong.

Protocol coalitions:

Visa aligns with OpenAI and Stripe's MPP; Mastercard with Google's Universal Commerce Protocol. OpenAI appears in both, a potential fault line.

Shared plumbing:

Both networks build on Cloudflare's Web Bot Auth at the CDN layer, then diverge on everything above it.

Stablecoin hedge:

Both Visa and Mastercard are simultaneously building stablecoin infrastructure, effectively hedging against the interchange disruption Citrini modeled.

Market reaction:

Visa dropped 4.5%, Mastercard 5.7%, AmEx 7.2% on a single research note explicitly labeled "not a prediction."

Foundation model:

Mastercard unveiled a generative AI model trained on billions of anonymized transactions, positioning it as fraud detection and commerce backend infrastructure.

Research Notes

Context Rot: How Increasing Input Tokens Impacts LLM Performance

Across 18 frontier models, Chroma's research team found universal performance degradation well before advertised context limits. Effective capacity runs roughly 60–70% of maximum. Counterintuitively, models handled shuffled text better than coherent text, implicating attention architecture itself.

How severe is the drop-off?

Every model tested degraded. A 200K-token window can start failing around 50K, quietly eroding long-running agent reliability.

Who validated this?

Chroma published the original findings in July 2025; independent replication and citation across multiple sources followed.

Improving LLM Function-Calling via Guided-Structured Reasoning Templates

Proposes structured reasoning templates that walk models through step-by-step function-call generation, tackling documented 20–27% accuracy drops caused by rigid JSON output constraints. Achieves 3–12% relative improvements over strong baselines while making tool-use decisions more interpretable.

What's the production payoff?

Templates reduce tool-use errors across model families, dropping into existing agent pipelines without requiring architectural changes.

Why is this urgent now?

As MCP multi-server environments grow tool counts, function-calling accuracy compounds downward. Reliability fixes can't wait.

Multi-Agent Deep Reinforcement Learning Under Constrained Communications (DG-MAPPO)

A fully distributed multi-agent RL framework that strips out centralized training dependencies via a Distributed Graph Attention Network. Agents learn to coordinate through peer-to-peer communication alone, outperforming centralized baselines on StarCraft II and Google Research Football.

What's genuinely novel?

Believed to be the first framework fully eliminating privileged centralized information from multi-agent coordination and training.

Where does it land practically?

Enterprise multi-agent deployments face constant team changes and shifting environments, precisely where centralized paradigms break.

Agentics 2.0: Logical Transduction Algebra for Agentic Data Workflows

March 2026 preprint formalizing each LLM call as a typed semantic transformation composable through algebraic operators. These execute as stateless asynchronous calls in parallel Map-Reduce architectures, opening a path toward formal verification of agent behavior.

What's the conceptual shift?

Agentic workflows become algebraically composable typed functions rather than prompt chains, making formal verification tractable.

Who should pay attention?

Anyone grappling with the gap between research prototypes and enterprise requirements for reliability and observability in production.

Research Notes

Context Rot: How Increasing Input Tokens Impacts LLM Performance

Across 18 frontier models, Chroma's research team found universal performance degradation well before advertised context limits. Effective capacity runs roughly 60–70% of maximum. Counterintuitively, models handled shuffled text better than coherent text, implicating attention architecture itself.

How severe is the drop-off?

Every model tested degraded. A 200K-token window can start failing around 50K, quietly eroding long-running agent reliability.

Who validated this?

Chroma published the original findings in July 2025; independent replication and citation across multiple sources followed.

Research Notes

Improving LLM Function-Calling via Guided-Structured Reasoning Templates

Proposes structured reasoning templates that walk models through step-by-step function-call generation, tackling documented 20–27% accuracy drops caused by rigid JSON output constraints. Achieves 3–12% relative improvements over strong baselines while making tool-use decisions more interpretable.

What's the production payoff?

Templates reduce tool-use errors across model families, dropping into existing agent pipelines without requiring architectural changes.

Why is this urgent now?

As MCP multi-server environments grow tool counts, function-calling accuracy compounds downward. Reliability fixes can't wait.

Research Notes

Multi-Agent Deep Reinforcement Learning Under Constrained Communications (DG-MAPPO)

A fully distributed multi-agent RL framework that strips out centralized training dependencies via a Distributed Graph Attention Network. Agents learn to coordinate through peer-to-peer communication alone, outperforming centralized baselines on StarCraft II and Google Research Football.

What's genuinely novel?

Believed to be the first framework fully eliminating privileged centralized information from multi-agent coordination and training.

Where does it land practically?

Enterprise multi-agent deployments face constant team changes and shifting environments, precisely where centralized paradigms break.

Research Notes

Agentics 2.0: Logical Transduction Algebra for Agentic Data Workflows

March 2026 preprint formalizing each LLM call as a typed semantic transformation composable through algebraic operators. These execute as stateless asynchronous calls in parallel Map-Reduce architectures, opening a path toward formal verification of agent behavior.

What's the conceptual shift?

Agentic workflows become algebraically composable typed functions rather than prompt chains, making formal verification tractable.

Who should pay attention?

Anyone grappling with the gap between research prototypes and enterprise requirements for reliability and observability in production.