The Rise of 10x Founders: AI, Experimentation, and the Future of Technology with Jeff Bussgang

Jeff Bussgang: Thanks for having me. I've had the privilege of being a serial entrepreneur, taking multiple companies public. These days, I’m probably best known as a professor at HBS, teaching the popular “Launching Tech Ventures” course. I’ve been involved in AI since the '’80s—yes, you can say that on camera—and most recently wrote The Experimentation Machine.

Brian Elliott: You only put out content when there’s a moment that matters. Why write this book now?

Jeff: Well, first of all, I’m most famous for being your professor! But seriously, we’re in a pivotal moment. Founders are realizing they can personally leverage AI tools to build highly productive and effective organizations in pursuit of product-market fit.

Early on, there was confusion—people thought, “If I’m not building an AI company, do I need to learn AI?” That moment has passed. Now, no matter your industry—tech, services, anything—you can use these tools to become what I call a 10x founder, similar to the mythical 10x developer.

Sid Pardeshi: As a 10x developer and founder myself, I loved the book’s focus on this shift—from adding more people to asking, “Where can I use AI to solve this?” If there’s an API and the task is in English, an AI agent can probably do it. At Blitzy, we’re helping enable that future—building and maintaining software autonomously with AI agents.

Jeff: Exactly. There’s a shift in mindset—every time there’s a job to be done, don’t first think of hiring a person. Ask if AI, or a human-AI combo, can do the job more effectively. I heard a founder describe hiring individuals who come with their own team of AI agents—like a designer or PM bringing their personalized toolset.

Sid: Each person becomes a micro-agency. That’s the future.

Brian: I heard Blitzy made it into the book as a late addition. True?

Jeff: Yes, your story was too compelling to leave out. You’ve used over 3,000 agents in product development, showing the power of specialized AI agents. Not one “uber-agent” doing everything, but a chain of specialists—one agent’s output becoming another’s input. Your process—from specs to code to testing and review—exemplifies this beautifully.

Brian: Are you seeing more companies like this at Flybridge?

Jeff: Definitely. One of our investment theses is around agent management. Think 8–12 years ahead: companies will use more AI agents than humans. We’re investing now in the management layer—authentication, orchestration, explainability. We saw a similar trajectory with Zest AI, which we backed 15 years ago to use ML in credit underwriting. It took time, but now it’s mainstream.

Sid: Coordination is the complex challenge. At Blitzy, we’ve built an engine that can dynamically recruit and manage specialized agents. The agents aren’t trained from scratch—they’re instructed with tools and context. The real work is estimating task complexity, deploying the right agents, and handling failures with robust guardrails.

Jeff: Exactly. That’s why agentic frameworks like Crew AI are so interesting—managing roles, hierarchies, and responsibilities. You’re unlocking task accuracy and reliability. I’m curious—what do you think the killer apps for agentic AI will be?

Sid: Software development is the obvious starting point, but we’ll see more—booking, payments, memos. Any task where actions span multiple domains. The real power comes when agents can “think” through problems, validate each other’s work, and collaborate intelligently. That’s where test-time reasoning comes in.

Jeff: Great point. Jensen from NVIDIA even highlighted this—pre-training, post-training, and test-time reasoning are now distinct optimization stages. OpenAI’s advances in reasoning show the potential for agents to handle more complex tasks reliably.

Brian: One question I keep coming back to—how will agent programming become accessible to non-technical users?

Jeff: That’s key. Just like low-code and no-code made software development accessible, agentic AI will need conversational interfaces. Programming an agent could feel like giving instructions to an employee—conversational, intuitive, and powerful.

Sid: And in a way, ChatGPT is already doing that. You give it goals, it reasons through tasks, and it performs.

Sid Pardeshi: When users work in environments they already understand—like developers writing docs or PMs drafting PRDs—the ideal system shouldn’t ask them to think about how many agents to use. It should decide for them. That’s optimal UX. Historically, agentic AI tools have demanded prompt engineering—a once-hot skill—but we’re now moving beyond that. Tools like ChatGPT and Blitzy reduce that need.

Brian Elliott: So you’re saying prompt engineering as a job might be on its way out?

Sid: Exactly. The future lies in no-code and low-code agent creation. Software is just layers of abstraction, and English is the ultimate abstraction. But what’s less obvious is the deep specialization needed in these agent systems to enable true AI-native workflows. Tools like LangChain or Crew AI still struggle when paired with general-purpose models like GPT or Claude.

Jeff Bussgang: Which points to the gap between where models are today and where they need to be to support true low-code AI agents. But Sid, you’ve solved part of this—what’s Blitzy doing with LangChain and this new wave of AI-native dev tooling?

Sid: Right—we built a proprietary RAG (retrieval augmented generation) system, specifically graph RAG. Since LLMs have a knowledge cutoff (usually Oct 2023), they miss out on new SDKs, APIs, or enterprise-specific data. We crawl leading API docs and support OpenAPI schemas and custom SDKs. This ensures the code we generate is always based on the latest documentation—no deprecated packages, no security holes. We’re launching this soon.

Jeff: That’s amazing. You’re essentially building self-learning, continuously updating systems—eliminating tech debt before it begins.

Brian: Right, the idea that your system will always have the freshest API knowledge and generate safe, updated code is huge.

Jeff: There’s a parallel here with a fascinating experiment in Japan—an AI agent is performing ML research and submitting papers for peer review. It’s not just executing tasks—it’s inventing. OpenAI talks about five levels of AI:

1. Chatbots

2. Reasoning

3. Agentic

4. Organizational agents

5. Researchers (AI that invents and discovers)

That’s where it gets wild—agents generating new knowledge autonomously.

Sid: And what’s driving this is data. We’ve moved past data scarcity—the new bottleneck is compute. But now companies like OpenAI and Blitzy are generating synthetic data or tapping proprietary data. OpenAI used reinforcement learning with human feedback (RLHF) to build GPT-4. They’re not just consuming data—they’re creating it.

Jeff: So now agents can generate, learn from, and improve based on their own experiences—what you might call post-training and test-time reasoning.

Sid: Exactly. In domains like software development, there’s a ground truth—you either get the right output or you don’t. We log every step of an agent’s journey—what we call “agent scrolls.” These scrolls are massive datasets that capture the entire decision process and can be used to train even better models. It’s a closed loop.

Jeff: It’s wild how fast this is moving. But there’s a gap—organizations can’t adapt at the speed these tools are evolving. That impedance mismatch between tool capability and enterprise readiness is real.

Brian: So, how should enterprises respond?

Jeff: That’s what I wrote about in The Experimentation Machine. My advice: run experiments—top-down and bottom-up. I spoke to a public company CEO who immediately tested Microsoft’s new agent framework with his CTO. At the same time, he encouraged his team to try grassroots projects. You need both approaches.

Sid: But that “pilot to production” gap is huge. Many startups hit $2–5M in revenue but stall because enterprise customers can’t scale adoption across all departments. They love the pilot, but the rollout is slow.

Jeff: Exactly. Enterprises must define clear criteria: if the pilot succeeds, how do we scale? 2025–2026 will be about getting out of the prototype ghetto and into full enterprise deployment. Otherwise, agile startups will outpace them.

Sid Pardeshi: What we’ve seen in successful deployments is that you don’t just jump at a pilot because it’s a logo. Instead, we slow down to get the right stakeholders involved.

Before they get enterprise usage, that’s when they’re most excited. But they may not have the buy-in for what’s commercially viable. So we define upfront what success means and what commercialization looks like if those success metrics are hit.

Jeff Bussgang: That’s such a critical insight. In enterprise sales, there are two types of power:

The power to say yes to a $100K pilot The power to say yes to a $1M deployment

Startups often get stuck in pilot purgatory because they don’t identify and involve the right enterprise-level stakeholders early enough.

You have to ask: “If we hit these success criteria, do we move forward? Or is there another gate to get through?”

Jeff Bussgang: Years ago, we wrote about "AAA" – not just Applied AI but Absorbable Applied AI.

Some organizations are good at running experiments but poor at enterprise-wide absorption. Startups need to coach customers on how to scale adoption.

Brian Elliott: I spoke with a Fortune 50 CTO recently. They were running 800 pilots and had adopted zero. That told us we shouldn’t pursue them right now – they don’t have absorption capabilities.

Sid Pardeshi: When we talk to CTOs, the most promising signal is if your product maps to one of their top 3 priorities.

Often, there’s tech debt blocking their roadmap – if you can help unblock that, even on a small scale, they’ll lean in.

The other blockers? Compliance, security posture, ease of integration, and whether your team can handle onboarding without taxing their team.

Jeff Bussgang: This goes back to a line from my book: “Timeless methods, timely tools.”

You must identify a must-have, hair-on-fire problem. It doesn’t matter how cutting-edge your tool is—if it’s not tied to a top-3 priority, it won’t get traction.

Brian Elliott: If you're dealing with source code, it’s hard to earn trust right away.

So what we’ve done is replicate the results on an open-source code base that parallels the customer’s use case.

That raises trust dramatically – from 40% to nearly 100%. You're decreasing perceived risk and showing results quickly

Jeff Bussgang: CTOs and CIOs are exhausted hearing the same pitch. That’s why Product-Led Growth (PLG) can be powerful.

For example, at MongoDB, developers adopted it bottoms-up. Eventually, the Goldman Sachs CIO called them because MongoDB was in use across 100 instances.

In AI, there’s tension between top-down trust and bottom-up adoption. Your go-to-market strategy depends on your product. If it touches source code, you probably need to go top-down.

Sid Pardeshi: We’ve adopted a hybrid model:

A free product slice (e.g. generate documentation or product code) lets users see value without touching IP. Later, we engage them for the full enterprise offering.

This shortens the sales cycle and builds trust.

Jeff Bussgang: In AI, new moats come from three areas:

Proprietary data Workflow depth (system of record) Human-computer interface

For example, our portfolio company Noetica serves law firms by ingesting their historical private debt transactions and organizing them by bank, sector, size, covenants, etc. That proprietary data creates lock-in.
So yes—go top-down if it gets you access to unique data. Sometimes you have to slow down to go fast.

Sid Pardeshi: Jeff, I’ve said this before—your HBS class was the most fun I’ve had academically. We started Blitzy in your LTV class.

The dream? One day, you write a case study about us, and we sit in your class as the protagonists.

Jeff Bussgang: My new book, The Experimentation Machine, is available now for pre-order at www.jfbussgang.com and will be live on Amazon next month.

Excited to hear feedback from founders who use the tools to become 10x Founders.