General-purpose AI agents are displacing specialized software at the application layer, with value migrating to data and trust. The new 'AI services' stack is built around Action (orchestration) and long-running task management, with checkpoints, resume, identity propagation and scoped credentials as first-class requirements. Hugging Face has standardized the vocabulary. The 'long-running agent' is the new primitive.

05 Layer, AI services, long running agents

Three structural shifts in how software gets built around AI. The "application layer" — the layer where most SaaS companies live — is being reframed as the place where value used to live. AI services are reshaping the stack with new persistence, orchestration and identity layers that did not exist in 2024. And "long-running agents" is the formal name for the runtime pattern that every serious agent framework in 2026 is racing to own.

What You Need to Know: McKinsey-style analysis argues that general-purpose AI agents are about to displace specialized software at the application layer, with value migrating to data and trust layers instead. Neontri's enterprise-agent stack framework identifies Action (orchestration) as Layer 4 of the modern agent platform, with long-running task management and multi-agent coordination as the hard problem. MongoDB's "Designing an Agentic Platform" walkthrough lays out the runtime requirements: state and persistence with checkpoints, resume and crash recovery, security and governance with identity propagation. The Hugging Face agent glossary has now standardized the terms — agent, harness, scaffold, sub-agent, rollout, reward — so the rest of the industry can stop arguing about vocabulary.

Why It Matters

If you are an application company and you have not asked "will AI agents still need us in three years," you have not done the strategic exercise that 2026 demands. The honest answer for most workflow tools is "probably not in the form we are shipping today."
The "AI services" layer is the new infrastructure. It is where the new default spend goes. If your offering does not have a clear role in that layer, you are the application being replaced.
Long-running agents are the runtime pattern. Checkpoints, resume, crash recovery, identity propagation — these are now first-class requirements, not bolt-ons. The frameworks that get this right ship the 2027 default.
The vocabulary is finally settling. "Harness" and "scaffold" used to be used interchangeably. They are not. Hugging Face's glossary makes the distinction: harness is execution-focused, scaffold is behavior-defining.
For builders, the "long-running agent" is the new "stateless web service" — the unit of work that the next decade of infrastructure gets designed around.

What Actually Happened

The Application Layer Is Being Repriced

The SSRN paper "How General-Purpose AI Agents Displace Specialized Software" is the most-cited academic framing of the moment. The argument: the value migration that defined the SaaS era (from infrastructure to application) is reversing. General-purpose agents, with access to data and tools, can replicate most of what specialized software does — and the "best-of-breed" decision migrates from the application layer to the data and trust layer. The companies that own data, integration and trust (think Bloomberg, FactSet, the financial-data oligopoly) keep the value. The companies that own workflow do not. (SSRN)

Research Affiliates' "The Impact of AI on SaaS" framework takes it from the investor side: as AI agents vertically integrate into the application layer and displace SaaS through APIs, the SaaS multiple compresses. The winners are the data-and-trust providers. The losers are the workflow tools that can be replicated by a general-purpose agent plus a connector. (Research Affiliates)

RTInsights' "Agentic AI and the Death of SaaS" is the punchier version of the same argument: "Agentic AI changes where value is created in the software stack. In the SaaS era, value resided in the application layer, which is the interface. In the agent era, value resides in the data and the tools." (RTInsights)

The AI Services Layer Is The New Stack

Neontri's "Enterprise AI Agents: 2026 Strategy & Deployment Guide" lays out the modern agent platform as a layered stack: Memory (vector DBs, knowledge graphs), Planning (reasoning engines, task decomposers), Tool Use (function calling, API gateways), and Action (orchestration) — the layer that "manages long-running tasks and multi-agent coordination." Options for the Action layer include LangGraph for stateful workflows, Temporal for durable execution, and a growing set of agent-runtime platforms. (Neontri)

MongoDB's "Designing an Agentic Platform" walkthrough is the more concrete reference architecture. The non-negotiable requirements for production agent infrastructure: state and persistence (checkpoints, resume, crash recovery for long-running agents); security and governance (identity propagation, audit logs, scoped credentials); orchestration (sub-agents, parallel task execution, retries with backoff); observability (per-step traces, cost attribution, prompt+completion logging). (MongoDB)

Northflank's "Top AI agent runtime tools and platforms in 2026" is the buyer's-guide version. The required features line up: "long-running stateful services for memory and agent state." Bring-your-own model. Bring-your-own tools. The runtime is the new platform. (Northflank)

The Long-Running Agent Is The New Primitive

Agentuity's "AI Agent Infrastructure: The Complete Guide" gets the framing right. "Agent infrastructure is a distinct layer with its own requirements: runtime support for long-running stateful processes, orchestration for multi-agent systems, durable execution." This is the layer that the LangGraph / Temporal / Inngest / Restate crowd is building. (Agentuity)

Stack Overflow's "AI agents will succeed because one tool is better than ten" is the architectural counter-argument. Zach Lloyd's framing: the agent that wins is the one with the deepest tool integration, not the most tools. The long-running agent needs to be able to do one thing end-to-end without offloading to a human. (Stack Overflow)

The Hugging Face "Harness, Scaffold, and the AI Agent Terms Worth Getting Right" glossary is the standardization play. An AI agent is the combination of a core model with a behavior-defining scaffold and an execution-focused harness. These systems use context engineering and sub-agents to manage memory and break down complex objectives into specialized, autonomous tasks. In training scenarios, reinforcement learning structures like rollouts and rewards provide the essential data and scoring mechanisms needed to refine model weights through environmental interaction. The vocabulary is now stable. (Hugging Face)

What Application Companies Should Do

The TLDR Product digest this week pointed at the same problem from a different angle. The "From SaaS Tools to AI Workspaces" thread captures the user-side version: if the unit of work becomes the agent, the SaaS tool you log into is the wrong layer. The companies that survive are the ones that ship a Claude Cowork plugin or an MCP app, not the ones that ship a SaaS dashboard. (Thread Reader)

For application companies, the three-year question is concrete. Will your customers describe your product as "the dashboard" or as "the agent's pantry"? The pantry companies (data, trust, identity, distribution) get to keep pricing power. The dashboard companies become implementation detail.

The Take

The "AI services" framing is the one I want builders to internalize. The layer where you sit in the stack is now a strategic choice, not a tactical one. If you are an application company, you are sitting in the layer that general-purpose agents are about to compress. The companies that survive are the ones that move up the stack — into data, into trust, into identity, into distribution — or sideways into infrastructure that the agents cannot replace.

The long-running agent is the unit of work that the 2027 default infrastructure is being designed around. Checkpoints, resume, durable execution, scoped identity — these are not optional. If you are picking an agent framework in 2026, the runtime's treatment of long-running state is the single most important feature. The vocabulary (harness, scaffold, sub-agent, rollout, reward) is now stable. Use it.

The hardest question in this digest is the one CFOs are going to start asking their portfolio companies: "What is your defensible layer in 2028?" If the answer is "the application," start having a different conversation.

Quick Summary

McKinsey-style analysis says general-purpose AI agents are displacing specialized software at the application layer, with value migrating to data and trust. The new "AI services" stack is built around Action (orchestration) and long-running task management, with MongoDB and Neontri laying out the production requirements: checkpoints, resume, identity propagation, scoped credentials. Hugging Face has standardized the vocabulary — agent, harness, scaffold, sub-agent, rollout, reward. The "long-running agent" is the new primitive.

05 Layer , AI services , long running agents