NFX's New Thesis on Defensibility in the Agent Era

The Defensibility Question That Keeps VCs Up at Night

For the past eighteen months, venture capital has poured into AI with almost religious fervor. According to data tracking the market, AI gets 31% of venture funds in Q2, Q3 2024, a staggering concentration that reflects the sector's dominance. Yet beneath the headlines about billion-dollar AI rounds and breakneck scaling lies a persistent anxiety: How do you build a defensible business when the underlying models are commoditizing at light speed?

NFX, the thesis-driven venture firm known for its sharp frameworks on network effects and founder psychology, has been quietly wrestling with this problem. The firm recently updated its foundational defensibility thesis-originally built for the mobile and SaaS eras-to account for a new reality: the rise of AI agents.

This isn't a minor tweak. It's a fundamental recalibration of what defensibility means when software can think, act, and learn autonomously. For founders raising capital in 2025, understanding NFX's updated framework is no longer optional. It's the language your investors are speaking.

What NFX's Original Defensibility Framework Actually Said

Before diving into the new thesis, you need to understand what came before. NFX built its reputation on identifying the few durable sources of competitive advantage in software. The firm's framework identified seven core defensibilities:

Network effects: The classic moat. Your product becomes more valuable as more people use it. Slack's network effects within teams, for instance, make it sticky and hard to displace.

Scale economies: The ability to serve more customers at lower marginal cost. Cloud infrastructure providers like AWS benefit from this relentlessly.

Brand: Emotional and rational associations that command premium pricing and loyalty. Vercel in the developer tools space has built brand weight that transcends its underlying technology.

Switching costs: The friction required to move to a competitor. Enterprise software often relies on this-the cost to migrate data and retraining is prohibitive.

Data: Proprietary datasets that improve product performance or enable insights competitors can't replicate. This was always the promise of AI, but the execution has been messier than anticipated.

Regulatory: Licenses, certifications, and compliance barriers that competitors can't easily hurdle. Healthcare and fintech startups often lean on this.

Proprietary technology: Patents, algorithms, or engineering breakthroughs that are genuinely hard to copy. This one has always been the trickiest to prove.

For years, this framework worked. A founder could raise a Series A by showing early traction in one or two of these categories. By Series B, you'd better have two or three locked in. By Series C, you were expected to have a defensibility moat so durable that large competitors couldn't easily replicate it.

Then AI agents arrived and scrambled the board.

Why AI Agents Break the Old Framework

An AI agent isn't just a chatbot or a prediction engine. It's a piece of software that can perceive its environment, set goals, take actions, and learn from feedback-often without explicit human instruction for each step. Think of it less like a calculator and more like an intern who can handle a task from start to finish.

As explored in depth in The AI Defensibility Playbook, the challenge with agents is that they collapse the time it takes for competitors to build similar functionality. A feature that once took six months to engineer-say, an autonomous customer support workflow-can now be assembled in weeks by combining a large language model with a few integrations and prompt engineering.

This creates a crisis for the traditional defensibility narrative. If anyone with access to GPT-4, Claude, or an open-source model can build a functionally similar agent in a matter of weeks, what's the moat? How do you justify a $100 million valuation when your core technology is a commodity?

NFX's insight is that the answer lies not in the agent itself, but in what the agent knows and what it can do that others can't. The defensibility in the agent era shifts from how you build to what you build with and where you deploy it.

The Three Pillars of NFX's Agent-Era Defensibility

NFX has reframed defensibility around three interlocking pillars, each of which founders need to prove out earlier than in previous cycles.

Pillar 1: Domain-Specific Data and Context

In the agent era, raw computational power is worthless without domain knowledge. An agent that can write code is interesting. An agent that understands your company's codebase, architecture patterns, and deployment practices is defensible.

This is where proprietary data becomes genuinely valuable again-but not in the way founders initially thought. It's not about hoarding customer data to train a better model (a strategy that's become legally and ethically fraught). It's about building agents that have deep, structured knowledge of a specific domain or vertical.

Consider a healthcare revenue cycle agent. The commodity part-the ability to read documents and extract information-is trivial now. Every LLM can do that. But an agent that understands healthcare billing codes, insurance coverage rules, prior authorization workflows, and the specific quirks of your hospital's EHR system? That's defensible. It's built on domain-specific context that takes months or years to accumulate and encode.

For founders, this means you need to prove early that your agent has access to proprietary or hard-to-replicate domain knowledge. This might come from:

• Exclusive partnerships: Agreements with industry incumbents that give you access to their data, workflows, or integration points. A manufacturing agent that integrates with Siemens equipment has a moat a generalist competitor can't easily replicate.

• Proprietary datasets: Information you've collected or licensed that competitors can't easily access. A legal tech agent trained on your firm's case outcomes and precedents is more defensible than one trained on public legal documents.

• Embedded workflows: Deep integration into how your customers actually work. An agent that sits inside Salesforce and understands your customer's CRM structure, custom fields, and sales process is stickier than one that sits outside and tries to sync data.

NFX expects founders to articulate this clearly in pitch meetings. Not "we have access to data"-that's table stakes. But "we have exclusive access to X domain knowledge that takes a competitor 18 months to replicate," backed by specific examples.

Pillar 2: Execution Moat and Reliability

As detailed in AI Agents: The Next Frontier In Enterprise Automation, the gap between a working prototype and a production-grade agent that enterprises trust with their workflows is enormous.

This is where many founders stumble. They build a proof-of-concept agent that works 80% of the time in a demo, then assume scaling to 99.9% reliability is just engineering work. It's not. It's a defensibility moat.

An agent that reliably handles exceptions, recovers from failures, and provides transparent reasoning about its decisions is worth 10x more than one that works most of the time. This is because the cost of agent failure in production is asymmetric. If your expense approval agent approves $50,000 in fraudulent expenses, the reputational and financial damage can be severe. This creates switching costs and lock-in that competitors can't easily overcome.

NFX now expects founders to demonstrate this early:

• Failure rate data: Not "we're working on reliability." But "our agent successfully completes 97% of transactions without human intervention, and we've documented the failure modes and recovery strategies."

• Audit trails and explainability: Enterprise customers increasingly demand to understand why an agent made a decision. Founders who've built robust logging, explainability, and audit capabilities have a genuine moat. A competitor can copy your agent's logic, but replicating your audit trail infrastructure takes months.

• Graceful degradation: When an agent is uncertain, does it escalate to a human? Does it ask clarifying questions? Does it fail silently? Founders who've thought deeply about these edge cases and built systems to handle them are further ahead.

This is why AI agents are finally having a moment - here's how to build one emphasizes that building defensible agents requires treating reliability as a first-class feature, not an afterthought.

Pillar 3: Customer Lock-In Through Integration Depth

The third pillar is perhaps the most concrete: how deeply your agent is woven into your customer's workflows.

A generic agent that sits on top of your customer's systems and sends recommendations is easy to replace. An agent that's integrated into their data pipelines, embedded in their tools, and has learned the specific quirks and preferences of their team is not.

This is the modern incarnation of switching costs. It's not about contractual lock-in (though that helps). It's about architectural lock-in. The deeper your agent is embedded, the more it learns about your customer, and the more valuable it becomes relative to alternatives.

NFX expects founders to think about this from day one:

• API-first architecture: Can your agent integrate deeply into your customer's existing tools? If you're building an agent for sales teams, does it live inside Salesforce, or does it sit outside and try to sync data? The former is more defensible.

• Continuous learning: Does your agent get better over time as it observes your customer's workflows? This is a genuine moat. A competitor's agent starts at zero; yours starts at the level of sophistication it's reached with your customer.

• Custom configurations: How much can your customer customize the agent's behavior without requiring engineering work from you? The more self-service customization you enable, the more your customer has invested in your specific platform.

Think of it this way: Slack's defensibility isn't just about network effects. It's also about the fact that every team has thousands of hours invested in configuring integrations, building workflows, and training people on Slack's specific interface. The switching cost is real.

What Founders Need to Prove Now, Not Later

Under NFX's updated framework, the timeline for proving defensibility has compressed. In the SaaS era, you could raise a Series A on the promise of defensibility. You'd show strong unit economics, retention, and a plausible path to a moat, and investors would wait for you to build it out.

In the agent era, that's no longer credible. The barrier to entry is too low. Anyone with $1,000 and a weekend can prototype an agent. VCs know this. So they're asking for proof earlier.

Here's what NFX and similar thesis-driven funds now expect to see at different stages:

At pre-seed and seed: Proof that you have a genuine domain insight that translates to a defensible agent. This might be:

• You've worked in the industry for a decade and understand workflows that no one outside the industry knows about.
• You have exclusive access to data or partnerships that give your agent unfair advantages.
• You've built a prototype that solves a specific, painful problem better than existing solutions-not because you're smarter, but because you understand the domain deeply.

The pitch should not be "we're building an AI agent for X." It should be "we're building an AI agent for X, and we have Y advantage that competitors can't easily replicate." And that advantage should be concrete, not aspirational.

At Series A: Proof that your agent is production-grade and locked into customers:

• Specific metrics on reliability, accuracy, and time-to-value.
• Evidence that customers are dependent on your agent (retention rates, usage depth, switching cost analysis).
• Clear documentation of what makes your agent better than a generic alternative, and why that advantage is durable.

For founders raising Series A, this is where AI Startup Valuations: The Reality Check You Need for Fundraising Success becomes essential reading. Valuations in the agent space are increasingly tied to defensibility metrics, not just user growth.

At Series B and beyond: Proof that you're building a defensible platform, not just a single agent:

• Can you apply your core defensibility (domain knowledge, execution excellence, integration depth) to adjacent use cases?
• Are you building network effects or other second-order defensibilities on top of your agent?
• What's your path to becoming the operating system for a category?

How This Changes Your Pitch Deck

If you're raising capital for an agent-native startup, here's how NFX's updated defensibility thesis should reshape your pitch:

Slide 1-2: Problem and Insight: Lead with a specific, domain-level insight that only you (or a very small group of people) have. Not "AI agents can automate customer support." But "customer support in fintech is broken because agents need to understand KYC rules, fraud patterns, and regulatory compliance-and we've built an agent that does."

Slide 3-4: Why You're Defensible: Explicitly articulate your defensibility thesis. Which of the three pillars (domain knowledge, execution excellence, integration depth) are you leading with? Why can't a competitor easily replicate it?

Slide 5-6: Proof Points: Show specific evidence. If domain knowledge is your moat, show your exclusive partnerships or proprietary datasets. If execution excellence is your moat, show reliability metrics and customer testimonials about uptime and accuracy. If integration depth is your moat, show usage metrics and switching cost analysis.

Slide 7-8: Roadmap: How will you expand your defensibility? Will you add new domain areas? Deepen integration? Build network effects? This is where investors evaluate whether you're building a durable business or a point solution.

For more tactical guidance on pitching AI startups specifically, A Step-by-Step Guide for Entrepreneurs on How to Pitch Their AI Projects and Raise Private Money provides a detailed playbook.

Real-World Examples: How the Framework Plays Out

Let's ground this in concrete examples. Consider three hypothetical agent-native startups, all raising Series A:

Startup A: Generic Customer Support Agent

They've built an agent that can handle common customer support tickets across industries. It integrates with Zendesk, Intercom, and Freshdesk. It's 85% accurate and handles 40% of inbound tickets without human intervention.

Under NFX's framework, this is a weak defensibility position. The technology is replicable. The integration is standard. There's no domain-specific knowledge that competitors can't quickly copy. Intercom or Zendesk could build this feature themselves. This startup would struggle to raise at a premium valuation.

Startup B: Healthcare Revenue Cycle Agent

They've built an agent specifically for healthcare revenue cycle management. It understands billing codes, insurance coverage rules, prior authorization workflows, and the specific EHR systems used in large hospital networks. They've partnered with three major hospital systems that use their agent to process claims. The agent achieves 94% accuracy on claim submissions, saving each hospital partner $2-3 million annually in denied claims.

This is a strong defensibility position. The domain knowledge is deep and hard to replicate. The partnerships create exclusive data access. The integration is tight (embedded in the hospital's workflows). Competitors would need domain expertise, partnerships, and integration work to compete. This startup would command a premium valuation.

Startup C: Manufacturing Quality Control Agent

They've built an agent that inspects manufactured goods for defects. It integrates with computer vision systems and manufacturing equipment. It's learned the specific defect patterns, quality standards, and production workflows of five large manufacturers. The agent has reduced defect rates by 18% and improved throughput by 12%.

This is also strong defensibility, but for different reasons. The continuous learning aspect (the agent gets better as it observes each manufacturer's specific patterns) creates lock-in. The integration depth is high. The domain knowledge is deep. Competitors would struggle to replicate the agent's performance without access to the same manufacturing environments.

Notice what's common across startups B and C: they're not competing on the agent's underlying technology. They're competing on domain knowledge, integration depth, and execution excellence. That's NFX's insight.

The Broader Shift in VC Thesis

NFX isn't alone in this recalibration. As AI Agents Are the Web3 of This AI Cycle notes, the venture industry is converging on a shared insight: the defensibility of agent-native startups lies not in the agent itself, but in the ecosystem around it.

This has implications for how you should think about your business:

Your moat is not your model. Language models are commoditizing. The moat is what you know and where you're deployed.

Speed matters, but durability matters more. You don't need to be first to market. You need to be first to build something so integrated and domain-specific that competitors can't easily catch up.

Your customers are your competitive advantage. The more you learn from your customers, the better your agent becomes, and the harder you are to displace. This is why customer feedback loops and continuous improvement are so critical.

You need to think like a platform company. Single-use agents are vulnerable. Platforms that can support multiple agent types, integrate deeply, and expand across use cases are more durable.

For founders raising capital, especially those looking at Series A and beyond, understanding this shift is critical. Investors like NFX are increasingly skeptical of agent startups that don't have a clear defensibility story. They've seen too many demos of impressive agents that could be replicated in a few weeks.

How to Build Defensibility Into Your Agent from Day One

If you're building an agent-native startup, here's how to bake defensibility into your product from the beginning:

Start with domain expertise, not technology. Don't ask "what can I build with AI agents?" Ask "what do I know deeply that others don't?" The domain expertise comes first. The agent is the tool you use to leverage that expertise.

Build proprietary datasets early. Start collecting and organizing domain-specific data from day one. This doesn't need to be customer data. It could be internal data about your industry, public data organized in a novel way, or exclusive partnerships that give you data access. The earlier you start, the larger your moat.

Design for integration and embedding. Don't build your agent as a standalone tool. Design it to integrate deeply into your customer's existing systems. This creates switching costs and lock-in.

Measure reliability obsessively. Track failure rates, accuracy, time-to-value, and customer satisfaction with granular precision. This is your proof of execution excellence. When you're raising capital, these metrics become your defensibility story.

Build for continuous learning. Design your agent to improve over time as it observes your customer's workflows. This creates a defensibility moat that gets stronger with time, not weaker.

Think about category creation. The strongest defensibility comes from owning a category. Think about how your agent can become the standard way a specific type of work gets done. This is where network effects and brand defensibility come in.

The Valuation Implications

Understanding NFX's defensibility framework has direct implications for how your startup gets valued.

In the SaaS era, valuations were often based on ARR multiples. A $5M ARR SaaS company might be valued at 10x revenue, or $50M. Simple.

In the agent era, valuations are increasingly based on defensibility metrics. A $5M ARR agent startup with weak defensibility might be valued at 3-5x revenue. The same $5M ARR with strong defensibility might be valued at 15-20x revenue.

The difference isn't the revenue. It's the moat. Investors are paying for durability and the ability to maintain margins as competition increases.

For a deep dive into how AI startup valuations work in practice, AI Startup Valuations: The Reality Check You Need for Fundraising Success breaks down the specific metrics and multiples investors are using in 2025.

What This Means for Different Founder Types

NFX's updated defensibility framework has different implications depending on your background and experience:

If you're a domain expert (former healthcare executive, manufacturing engineer, etc.): You're in an excellent position. Your domain expertise is your primary defensibility. Lean into it. Build your agent around the specific insights and workflows you understand. Your pitch should lead with "I spent 15 years in this industry, and I understand problems that AI agents can solve better than anyone else."

If you're a technologist (ML engineer, software engineer): You need a domain partner or deep domain knowledge to be competitive. Pure technology is not defensible. Find a co-founder or advisor with domain expertise. Or pick a vertical and become an expert through customer conversations and research. Your pitch should not lead with technology. It should lead with domain insight.

If you're a product person: You're in the middle. You can build great agents, but defensibility comes from understanding customer workflows and building deep integration. Focus on customer intimacy and understanding the specific workflows you're automating. Your pitch should lead with customer insights and evidence of lock-in.

If you're a first-time founder: You have the hardest job. You need to either have domain expertise, find co-founders who do, or spend 6-12 months building deep customer relationships in a specific vertical before raising capital. Don't try to raise on a general-purpose agent. It won't work.

The Competitive Landscape: Who's Winning Under This Framework

Looking at agent-native startups that have raised significant capital in the past 12 months, the ones that align with NFX's defensibility framework are winning:

Startups with exclusive partnerships or data access: Companies that have locked in partnerships with major incumbents or have exclusive data access are raising at premium valuations. They've solved the defensibility problem.

Startups with deep vertical focus: Companies building agents for specific industries (healthcare, legal, manufacturing, finance) are raising at better terms than horizontal players. They have domain defensibility.

Startups with proven execution excellence: Companies that can demonstrate high reliability, accuracy, and customer satisfaction are commanding premium valuations. Execution excellence is defensibility.

Startups with tight customer integration: Companies that have embedded their agents deeply into customer workflows are seeing better retention and expansion. Integration depth is defensibility.

Notice what's not on this list: startups with impressive technology but no clear defensibility. They're struggling to raise or raising at down rounds.

Looking Ahead: How This Framework Will Evolve

NFX's updated defensibility framework is not the final word. As the agent market matures, the framework will continue to evolve. Here's what we might expect:

Increased emphasis on multi-agent systems: As agents become more capable, the defensibility will shift from single-agent performance to orchestration and coordination across multiple agents. The startup that can manage a fleet of specialized agents will have a moat.

Network effects becoming more relevant: As agents become more prevalent, network effects will matter more. An agent that learns from a network of similar agents (while respecting privacy) will be more defensible than one that learns only from its own customers.

Regulatory defensibility: As regulators catch up to agent technology, compliance and regulatory expertise will become a defensibility moat. Startups that can navigate the regulatory landscape will have an advantage.

Brand and trust: As agents handle more critical workflows, brand trust will become increasingly important. A startup with a strong reputation for reliability and ethics will command a premium.

For founders thinking about long-term defensibility, these are the trends to watch.

Practical Next Steps for Founders

If you're raising capital for an agent-native startup, here's what to do immediately:

1. Articulate your defensibility thesis: Write down, in one paragraph, why your agent is defensible. Which of the three pillars (domain knowledge, execution excellence, integration depth) are you leading with? Why can't a competitor easily replicate it? If you can't answer this clearly, you're not ready to raise.

2. Gather defensibility proof points: If domain knowledge is your moat, document your exclusive partnerships or proprietary datasets. If execution excellence is your moat, measure and document your reliability metrics. If integration depth is your moat, show customer usage and switching cost analysis. You need concrete evidence.

3. Benchmark against competitors: Who else is building in your space? How is your defensibility different? What would a competitor need to do to match you? If the answer is "not much," you need to strengthen your moat before raising.

4. Map your defensibility roadmap: How will you strengthen your moat over the next 18-24 months? Will you deepen domain knowledge? Improve execution excellence? Expand integration? Investors want to see that you're thinking about long-term defensibility, not just short-term traction.

5. Tailor your pitch to defensibility: Your Series A pitch should lead with defensibility, not technology. Spend 40% of your pitch on why you're defensible. Spend 20% on the market opportunity. Spend 20% on traction. Spend 20% on team. This allocation reflects what investors actually care about.

For a step-by-step guide on pitching AI projects specifically, A Step-by-Step Guide for Entrepreneurs on How to Pitch Their AI Projects and Raise Private Money walks through the entire process with examples and templates.

The Bottom Line

NFX's updated defensibility framework for the agent era is a reality check for founders. The days of raising capital on impressive technology demos are over. Investors have seen too many demos. They're now asking: Why can't someone else build this in a few weeks?

The answer has to be: Because we have domain knowledge they don't. Because our execution is so reliable that switching costs are too high. Because we're embedded so deeply in our customers' workflows that we're indispensable.

These are defensibility moats for the agent era. And they're not optional. They're table stakes for raising capital in 2025.

For founders building agent-native startups, the message is clear: Start with domain expertise, not technology. Build for integration and lock-in. Measure execution excellence obsessively. And tell a clear story about why your defensibility is durable.

That's the thesis. That's what investors like NFX are looking for. And that's what will separate the winners from the also-rans in the agent era.

For more context on how the broader AI investment landscape is evolving, AI Gets 31% of Venture Funds in Q2, Q3 2024: A Deep Dive into the VC Landscape provides market-level data on where capital is flowing and why. And if you're specifically interested in how to position your AI startup for success, 10 Game-Changing AI Startup Ideas That Will Skyrocket Your Valuation and Attract Investors explores which AI verticals and use cases investors are most excited about in 2025.

The agent era is here. The defensibility rules have changed. Now it's up to you to build accordingly.