For twenty years, enterprise software looked roughly the same.

Human → UI → Software → Data

Software provided tools. Humans executed workflows.

AI agents flip this model.

The emerging architecture

Human
  ↓
Agent Platform
  ↓
Agents
  ↓
Skills
  ↓
Enterprise Systems

Humans provide intent. Agents execute workflows.

1. The Agent Platform

The runtime for agents—analogous to what operating systems were for applications.

Responsibilities: running agents, routing tool calls, scheduling jobs, managing permissions, coordinating multiple agents.

Without this layer, you only have a single assistant. With it, you can run dozens of specialized agents in parallel.

2. Agents as roles, not features

Agents represent roles—Sales, Support, Finance, Document Processing—not individual product features.

Each agent has a defined goal, a set of skills, memory of past work, and bounded permissions to certain systems.

A Customer Intake Agent might read emails, extract documents, create CRM records, and schedule meetings—end to end, without a human in the loop. A digital employee responsible for a workflow.

3. Skills: where agents become actors

Agents don't execute actions directly. They call skills.

send_email()
crm.create_lead()
pdf.extract_text()
browser.fill_form()

Skills are structured, versioned wrappers around APIs and system operations. Without them, agents can reason. With them, agents can act.

4. Memory: the difference between an assistant and a collaborator

Memory allows agents to carry context across sessions: customer profiles, document histories, past conversations, company knowledge.

Short-term assistants don't change how companies work. Persistent agents do.

5. Enterprise systems: same systems, new operators

CRM, ERP, accounting software—these don't disappear. They move down the stack.

Agent → Skill → CRM API

The system remains the system of record. Agents become the operators. The enterprise doesn't replace its software investments—it wraps them.

What changes

In the SaaS era, employees learned software. In the agent era, software learns workflows.

A typical stack: one agent platform, 10–50 specialized agents, hundreds of skills, existing SaaS systems underneath.

Employees stop navigating software. They describe goals.

Where the hard work actually is

Most businesses don't need a general AI assistant. They need agents that understand their specific workflows.

The difficult part isn't the model. It's the system around the model: the skill library, the memory architecture, the permission model, the observability stack.

This is where most AI implementations stall—not at the model level, but at the operational level.

Where value accumulates

The most durable advantage may not come from the platform layer.

It may come from industry-specific skills and encoded workflows: insurance underwriting, immigration case processing, financial document interpretation. Once those are captured and tested in production, agents can execute them at scale.

That's where automation becomes real.

The shift

Enterprise software is moving from:

Human → Software

to:

Human → Agents → Software

This isn't a new UI. It's a new operating model for work.

The question for most companies is no longer whether to build toward this model. It's how fast, and starting where.