CURRENT | Market Pulse

The Runtime Layer

Where the Dependency Forms

By Nora Kaplan— April 29, 2026

Feature image for article: Where the Dependency Forms

When AWS customers reportedly greeted OpenAI's arrival on Bedrock with a shrug, the obvious read was that OpenAI had lost its edge. These were teams that had spent years building production systems on Claude, Llama, Mistral. They'd learned that models swap out in an afternoon.

The deal structure tells a different story, though. OpenAI gave Azure exclusive access to its APIs. It gave AWS something different: exclusive rights to co-developed infrastructure it hasn't offered anyone else, involving agent identity, persistent memory, and enterprise security boundaries. Two clouds, two very different assets. The split reveals a quiet bet about where real dependency actually forms.

The Runtime Layer

Where the Dependency Forms

When AWS customers reportedly greeted OpenAI's arrival on Bedrock with a shrug, the obvious read was that OpenAI had lost its edge. These were teams that had spent years building production systems on Claude, Llama, Mistral. They'd learned that models swap out in an afternoon.

The deal structure tells a different story, though. OpenAI gave Azure exclusive access to its APIs. It gave AWS something different: exclusive rights to co-developed infrastructure it hasn't offered anyone else, involving agent identity, persistent memory, and enterprise security boundaries. Two clouds, two very different assets. The split reveals a quiet bet about where real dependency actually forms.

Nora Kaplan

Nora Kaplan, former collaboration platform product leader turned technology writer. Studied human-computer interaction and spent years designing tools for knowledge work. Now writes about AI agents, work transformation, and how enterprise software reshapes human capability at TinyFish.

Platform Moves

The spotlight stays fixed on frontier models and the big distribution deals. Fair enough. But one layer down, something arguably more consequential is taking shape: the runtime and infrastructure that agents will actually run on.

This batch of moves shares a pattern worth watching. Major platforms and service providers are restructuring how enterprise software, IT operations, and the open web will work when agents are first-class participants. SaaS platforms are exposing themselves as callable infrastructure. IT service providers are naming the readiness gap out loud. Identity standards are being proposed for an internet where the visitor might not be human.

None of it is settled. But the plumbing is going in, and plumbing has a way of determining what gets built on top.

Platform Moves

The spotlight stays fixed on frontier models and the big distribution deals. Fair enough. But one layer down, something arguably more consequential is taking shape: the runtime and infrastructure that agents will actually run on.

This batch of moves shares a pattern worth watching. Major platforms and service providers are restructuring how enterprise software, IT operations, and the open web will work when agents are first-class participants. SaaS platforms are exposing themselves as callable infrastructure. IT service providers are naming the readiness gap out loud. Identity standards are being proposed for an internet where the visitor might not be human.

None of it is settled. But the plumbing is going in, and plumbing has a way of determining what gets built on top.

Platform Shift

AWS Turns Its Contact Center Into Four Agent Suites

Amazon Connect quietly grew from a single contact-center product into four agentic AI solutions covering supply chains, hiring, healthcare, and CX. AWS is now building first-party agent applications on its own runtime, drawing on Amazon's actual logistics, hiring, and clinical workflows. These aren't demos. They're operational lineage turned into product.

SaaS Architecture

Adobe Decomposes Its Platform Into Agent Infrastructure

Adobe CX Enterprise replaces Experience Cloud with an agent-first architecture, exposing MCP servers and agentic skills to external orchestrators. Adobe's customer is now also any agent that needs its data and decisioning capabilities. Partnerships span AWS, Anthropic, Google Cloud, Microsoft, and OpenAI. Coherent strategy, though adoption timelines remain open.

Infrastructure Operations

NTT DATA Deploys Multi-Agent Enterprise Infrastructure Management

NTT DATA's SDI Services Agent orchestrates specialized agents across networking, data centers, cybersecurity, and workplace environments. The design choice worth noting: agents reason through live telemetry and policy guardrails to take deterministic action across multivendor estates. They act on problems, not just surface alerts for humans to triage.

Readiness Gap

Kyndryl Names the Enterprise Agentic Infrastructure Gap

The world's largest IT infrastructure services provider launched a maturity model and implementation blueprints built around a blunt admission: today's enterprise environments cannot reliably support agentic AI. Their own research finds nearly half of organizations investing heavily in AI struggle to achieve meaningful returns. The diagnostic is more interesting than the product. You cannot scale agentic workflows atop manual operating models.

Agent Identity

Cloudflare and GoDaddy Propose Open Agent Identity

Two companies with enormous structural web coverage are jointly proposing identity standards for an agentic internet. Cloudflare proxies roughly 20% of websites; GoDaddy hosts millions more. Their Agent Name Service provides DNS-like verification for agents, while Web Bot Auth adds cryptographic signatures. The ambition goes well beyond blocking bots. They want agents to transact as verified participants.

Research Radar

TrajAD: Trajectory Anomaly Detection for Trustworthy LLM Agents

A fine-tuned Qwen3-4B verifier that detects and pinpoints procedural errors mid-trajectory, enabling surgical rollback-and-retry instead of restarting entire tasks. Includes TrajBench, the first dataset built specifically for trajectory anomaly localization.

Why not just use a bigger model?

General-purpose LLMs consistently fail at step-level error localization. Process reliability demands specialized fine-tuning, not scale.

What shifts here?

Evaluation moves from "did the agent get the right answer" to "did the agent reason correctly along the way."

TriCEGAR: A Trace-Driven Abstraction Mechanism for Agentic AI

Learns formal state abstractions directly from agent execution traces using predicate trees and counterexample-guided refinement. Constructs a behavioral MDP online for probabilistic model checking, removing the manual state definitions that have kept formal verification out of production.

What's the bottleneck it targets?

Hand-authored state definitions that break every time agents or tool interfaces change. Learned predicates update automatically.

Where does this land in practice?

Probabilistic assurance queries and anomaly-based guardrails derived from execution logs, no developer-written specs required.

Benchmarking Agents in Insurance Underwriting Environments

Snorkel AI and credentialed underwriters built a multi-turn benchmark testing 13 frontier models against real enterprise friction: noisy tools, proprietary domain knowledge, imperfect simulated users. Pass^k evaluation exposes up to 20% performance degradation from research scores to production conditions.

What breaks under pressure?

Models hallucinate domain knowledge even with tool access. The most accurate models aren't the most efficient under realistic constraints.

Why does the benchmark design matter?

Expert-authored tasks with authentic noise finally measure the gap between demo performance and deployable reliability.

Ontology-Constrained Neural Reasoning in Enterprise Agentic Systems

Formal ontologies constrain LLM reasoning at the input level across role, domain, and interaction layers. Tested across 600 runs in five regulated industries, the architecture reveals an "inverse parametric knowledge effect": ontological grounding helps most where model knowledge is thinnest.

What's counterintuitive?

Ontological constraints become more valuable in specialized domains as models scale, not less. Weak parametric knowledge amplifies the benefit.

How real is the deployment?

A production system spanning 21 verticals with 650+ agents, showing measurable gains in regulatory compliance and role consistency.

Research Radar

TrajAD: Trajectory Anomaly Detection for Trustworthy LLM Agents

A fine-tuned Qwen3-4B verifier that detects and pinpoints procedural errors mid-trajectory, enabling surgical rollback-and-retry instead of restarting entire tasks. Includes TrajBench, the first dataset built specifically for trajectory anomaly localization.

Why not just use a bigger model?

General-purpose LLMs consistently fail at step-level error localization. Process reliability demands specialized fine-tuning, not scale.

What shifts here?

Evaluation moves from "did the agent get the right answer" to "did the agent reason correctly along the way."

Research Radar

TriCEGAR: A Trace-Driven Abstraction Mechanism for Agentic AI

Learns formal state abstractions directly from agent execution traces using predicate trees and counterexample-guided refinement. Constructs a behavioral MDP online for probabilistic model checking, removing the manual state definitions that have kept formal verification out of production.

What's the bottleneck it targets?

Hand-authored state definitions that break every time agents or tool interfaces change. Learned predicates update automatically.

Where does this land in practice?

Probabilistic assurance queries and anomaly-based guardrails derived from execution logs, no developer-written specs required.

Research Radar

Benchmarking Agents in Insurance Underwriting Environments

Snorkel AI and credentialed underwriters built a multi-turn benchmark testing 13 frontier models against real enterprise friction: noisy tools, proprietary domain knowledge, imperfect simulated users. Pass^k evaluation exposes up to 20% performance degradation from research scores to production conditions.

What breaks under pressure?

Models hallucinate domain knowledge even with tool access. The most accurate models aren't the most efficient under realistic constraints.

Why does the benchmark design matter?

Expert-authored tasks with authentic noise finally measure the gap between demo performance and deployable reliability.

Research Radar

Ontology-Constrained Neural Reasoning in Enterprise Agentic Systems

Formal ontologies constrain LLM reasoning at the input level across role, domain, and interaction layers. Tested across 600 runs in five regulated industries, the architecture reveals an "inverse parametric knowledge effect": ontological grounding helps most where model knowledge is thinnest.

What's counterintuitive?

Ontological constraints become more valuable in specialized domains as models scale, not less. Weak parametric knowledge amplifies the benefit.

How real is the deployment?

A production system spanning 21 verticals with 650+ agents, showing measurable gains in regulatory compliance and role consistency.

The Payment Rail

Visa's Intelligent Commerce Connect Positions the Payment Rail as Persistent Agent Infrastructure

Visa's Intelligent Commerce Connect, announced April 8, supports four agent payment protocols through a single integration point and accepts non-Visa cards. The openness is genuine. But the structural parallel to cloud runtimes is worth sitting with. Both layers advertise protocol agnosticism while quietly accumulating state that agents depend on: tokenized credentials and spend controls for Visa, task context and governance logs for cloud providers.

Visa predicts millions of consumers will use agents for purchases by the 2026 holiday season. The lock-in won't live in the protocols. It'll live in the state.

The Payment Rail

Visa's Intelligent Commerce Connect Positions the Payment Rail as Persistent Agent Infrastructure

Visa's Intelligent Commerce Connect, announced April 8, supports four agent payment protocols through a single integration point and accepts non-Visa cards. The openness is genuine. But the structural parallel to cloud runtimes is worth sitting with. Both layers advertise protocol agnosticism while quietly accumulating state that agents depend on: tokenized credentials and spend controls for Visa, task context and governance logs for cloud providers.

Visa predicts millions of consumers will use agents for purchases by the 2026 holiday season. The lock-in won't live in the protocols. It'll live in the state.

IN SUMMARY

Protocol scope:

ICC routes through Trusted Agent Protocol, Machine Payments Protocol, Agentic Commerce Protocol, and Universal Commerce Protocol via a single integration point

State accumulation:

Agent-specific tokens, commerce signals for disputes, and spend controls validated against original user instructions create compounding platform dependency over time

MCP integration:

Visa launched a Model Context Protocol server connecting agents to its API portfolio, plus an Acceptance Agent Toolkit enabling plain-language interaction

Cloud parallel:

The AWS Bedrock AgentCore integration puts both infrastructure layers holding state simultaneously, neither visible to nor portable into the other

Pilot scope:

Currently testing with AWS, Highnote, Mesh, and Payabli, building on over 100 partner collaborations and live agent transactions from late 2025