Agents are Eating the Software World

"Software is eating the world, but AI is going to eat software," declared Jensen Huang, NVIDIA's CEO, updating Marc Andreessen's famous 2011 prediction for the AI age.

Today that prediction is becoming reality faster than anyone anticipated.

The traditional OS layer has served us well for decades, looking something like this:

Software Applications
Application Layer
Middleware
Hardware/Firmware
Network Layer

But we are witnessing a fundamental architectural shift: Agents are inserting themselves as the new OS layer, fundamentally changing how humans interact with all the layers beneath.

Agents aren't just glorified AI-driven APIs, though the distinction matters. Think of an agent as software that can follow complex instructions and independently complete multi-step jobs that previously required human coordination.

Consider travel booking: researching deals, reading reviews, coordinating schedules across multiple people, managing preferences, making reservations, and handling contingencies. That's hours of cognitive work involving dozens of decisions. Now imagine an agent that orchestrates all of this—checking with your friends about dates, cross-referencing everyone's preferences, calling restaurants for reservations, and presenting you with final options or booking everything autonomously.

That's the difference between a tool and an agent: tools require you to operate them; agents operate on your behalf.

When agents become the primary interface between humans and software, the entire stack transforms:

Natural Language Interface (Human ↔ Agent)
Agent Orchestration Layer ← [THE NEW OS]
Traditional Software Applications
Application Layer
Middleware
Hardware/Firmware
Network Layer

The agent layer handles workflow orchestration, decision-making, and cross-application coordination that previously required human intervention. It's not just automation—it's intelligent orchestration that adapts to context and learns from outcomes.

The funding data tells a compelling story: Investors have poured around $700 million this year into seed rounds for AI agent companies, with 24 US AI startups raising $100M or more in 2025 alone.

More telling are the revenue efficiency metrics: Mercor ($4.5M revenue per employee) and Cursor ($3.2M per employee) already surpass Microsoft ($1.8M per employee) and Meta ($2.2M per employee). When agent-first companies are generating 2-3x the revenue per employee of traditional software giants, we're seeing a fundamental productivity shift.

Where traditional software required learning specific UIs, agents use natural language as the universal interface. No more clicking through menus or remembering keyboard shortcuts—just describe what you want accomplished.

Instead of manually switching between applications and managing complex workflows, agents coordinate across multiple systems according to predetermined instructions. It's like having a digital chief of staff that amplifies your daily progress and task completion.

Is this just "glorified RPA"?

Not quite. While Robotic Process Automation follows rigid scripts, agents adapt to context, handle exceptions, and improve through interaction. The difference is intelligence and autonomy.

• Cursor: AI-first code editor generating $3.2M revenue per employee

• Devin by Cognition Labs: Autonomous software engineering agent that can build entire applications

• GitHub Copilot: Transforming development workflows for millions of developers

• Mercor: AI recruitment agent achieving $4.5M revenue per employee by automating candidate sourcing and screening

• Customer Service Agents: Companies like Ada replacing traditional ticketing systems with conversational AI

• Sales Automation: Platforms like Outreach's AI SDR handling lead qualification and initial outreach

Recent moves by major players signal the infrastructure shift:

• Databricks acquiring Tecton to boost AI agent offerings

• OpenAI and Amazon Web Services making significant AI agent announcements

• Microsoft: Copilot agents across Office suite

• Google: Bard and Duet AI agents

• Amazon: Alexa for Business and AWS agent services

• OpenAI: GPT-based agent frameworks

The startup ecosystem is where the most interesting agent development is happening:

• Vertical Specialists: Industry-specific agents for healthcare, legal, finance

• Horizontal Platforms: Agent orchestration and development tools

• Infrastructure Plays: Tools for building, deploying, and managing agent workflows

Projects like AutoGPT, LangChain, and CrewAI are democratizing agent development, creating a community-driven innovation layer.

Forward-thinking enterprises are building internal agent capabilities, with early adopters seeing 40-60% productivity gains in knowledge work scenarios.

As I covered in my blog post on AaaS earlier this year, we're witnessing a fundamental shift from Software-as-a-Service to Agents-as-a-Service (AaaS). Instead of paying for software licenses that sit unused, companies pay for outcomes delivered by agents.

This shift enables new pricing models:

• Outcome-based pricing: Pay per completed task or achieved result

• Usage-based scaling: Costs align with actual agent activity

• Performance premiums: Better agents command higher prices through demonstrated ROI

Jensen Huang notes that progress in agentic AI is "spectacular and surprising," moving "faster and faster and getting into the flywheel zone."

When agents make decisions, understanding their reasoning becomes crucial for trust and debugging. The "black box" problem becomes more critical when agents take actions on your behalf.

How do you measure agent performance? Traditional software metrics (uptime, response time) don't capture agent effectiveness. New evaluation frameworks focusing on task completion, accuracy, and user satisfaction are emerging.

Agents must work seamlessly with existing enterprise systems, APIs, and workflows. The integration challenge is substantial, especially for large organizations with legacy infrastructure.

Agents will handle routine cognitive tasks—research, analysis, communication, scheduling—allowing humans to focus on strategy, creativity, and relationship-building.

Rather than replacing creativity, agents will amplify creative output by handling production tasks, research, iteration, and optimization.

The infrastructure supporting agents will become as critical as cloud infrastructure is today. Agent orchestration platforms, evaluation systems, and security frameworks will become essential enterprise infrastructure.

Winners will be those who blend:

1. Domain-specific intelligence: Deep expertise in particular industries or use cases

2. Infrastructure control: Robust platforms for agent deployment and management

3. Data advantages: Proprietary datasets that improve agent performance

4. Network effects: Agents that get better as more users interact with them

At risk are companies that:

• Treat agents as a feature rather than a fundamental architecture shift

• Underestimate the integration and reliability challenges

• Focus on technology without solving real workflow problems

• Ignore the new competitive dynamics of agent-vs-agent performance

We're not just seeing new applications—we're witnessing the emergence of a new operating system layer. Just as mobile OS platforms (iOS, Android) became the foundation for entire ecosystems, agent platforms will become the foundation for how humans interact with all digital services.

The companies that understand this shift and build accordingly will define the next decade of technology. Those that view agents as incremental improvements to existing software will find themselves disrupted by fundamentally different approaches to solving human problems.

The question isn't whether agents will eat the software world—it's whether you'll be building the agents or being eaten by them.

The views expressed are those of the author and do not necessarily reflect the views of any investment firm or portfolio company.