Founder Spotlight: Joe Hewett @Asteroid AI and bumper lanes for AI agents
Why there will soon be millions of AI agents roaming online and tools to build your own micro-products as an early-stage startup founder
Listen to a 10-minute pod on this article if you are on the go or just want a fun way to digest!
Every day, we find ourselves reflecting on the future of company building in the age of AI. One question keeps resurfacing in our conversations: What does the optimal AI-native tech stack look like for an early-stage startup across product, go-to-market, and back office?
Take it further—how does this manifest at the tactical level for a core founding team with only a few deeply committed individuals, no product, and no customers? How do you build something from nothing with nothing but bold ideas and cutting-edge AI at your disposal?
These are the puzzles that inspire us. And as we explore the countless AI tools in the open- and closed-source communities, we keep coming back to a core principle: AI-native founders don’t just use tools—they build their own.
Yes, starting with tools like ChatGPT, Claude, Perplexity, or specialized SaaS platforms like Copy.ai or Clay is great. They’re powerful, efficient, and accessible. But they’re just the tip of the iceberg. Being truly AI-native means leveraging AI to create personalized, internal micro-products that seamlessly integrate into your workflows. In other words, writing your own code.
Why? Three reasons:
Custom Needs: Every startup is unique, with micro, repetitive workflows that can’t always be addressed by off-the-shelf tools.
Cost Efficiency: Relying on pre-built tools can be expensive as your company scales.
Moat Creation: Building internal AI-powered systems creates an organizational edge that’s hard to replicate.
Imagine a co-founding team of two managing workflows that would typically require a 10-person team, from triaging thousands of job applications to iterating on customer feedback, all powered by agentic AI systems. A simple workflow, built with tools like LangChain, could transform how startups function at their core.
Our Approach
We practice what we preach. Over the last few months, we’ve built custom in-house micro-products to automate our own workflows. And while every founder’s needs will differ, here are the tools we recommend to code and build fast—no advanced programming skills required:
ChatGPT o1: Your go-to CPTO. Use it to map out high-level project structures, from backend and frontend to databases and APIs.
StackBlitz Bolt: Think of this as your Canva for frontend design. Rapidly visualize and iterate on product UIs with ease.
Cursor: Dive deeper into code with detailed prompts. Cursor is your Figma for customized product builds, with a learning curve shortened by its in-built AI.
Pro tip: Combine these tools iteratively. ChatGPT for strategy, Bolt for prototyping, and Cursor for fine-tuning. Last week, we built a WhatsApp UI clone in React and Tailwind CSS in under a minute using Bolt, which is mind-blowing.
No excuses, no limits. Build your workflows. Own your code. Your future moat depends on it.
In our last founder spotlight, we explored how AI agents are revolutionizing enterprise accounting. This time, we’re diving into the guardrails required to keep agentic AI systems safe and reliable once they’re out roaming the internet—thinking, planning, and taking actions on your behalf.
Enter Joe Hewett—co-founder & CTO of Asteroid. Joe envisions a near future where millions of AI agents operate online, and some inevitably go “rogue.” But thanks to startups like Asteroid, this dystopian possibility doesn’t have to become a reality. We sat with Joe at Entrepreneur First in London—right before his acceptance call into the Y Combinator Winter ‘25 batch (congrats Joe & team!)—and had an amazing conversation on how an AI-native future can be kept safe, secure, and thriving—at scale.
Let’s dive in:
“If you’re an AI agent, the world is your oyster…”
Pamir: Let’s start with your roots, Joe. Who are you, and how did you become a founder? What sparked your passion?
Joe: When I was 5 my brother showed me a drawing of Father Christmas on Microsoft Paint and it was over. I remember the first time a teacher executed a few lines of HTML and a webpage appeared and it blew my mind. I had always wondered how web pages were built. I generally enjoy tinkering with things. I was probably better at dismantling things than reassembling them though… All the way through school it was always obvious to me that I needed to study computer science. It was really hard to understand why there were only 4 students in my computing class. It seemed really obvious that it was going to be radically important for the coming decades. At 15, I was building websites for small local businesses and automating their workflows with Excel and Google Sheets scripts, and at 18 I started working in software engineering. I was never the best software engineer and the place that I worked had a ridiculous concentration of the world’s best computer scientists, but being the least skilled in the room is a good place to be. I started learning so much so quickly, and I had this realization that I wanted to spend my life building systems like them.
Pamir: Tell us about your previous work before Asteroid. How did it prepare you for founding your own company?
Joe: My prior experience at a cybersecurity company was unbelievably transformative. The organization encouraged innovation—the founder was especially forward-thinking and we worked on new and useful projects all the time. We had huge levels of autonomy and it was very much a startup despite existing for four decades. But as with many large organizations, eventually bureaucracy crept in and slowed things down. I wanted to try my hand at hard problems that I thought were most important–for the coming century–and be in an environment where I could build faster and with more autonomy.
The main project I worked on before starting Asteroid was using early AI systems to "scam the scammers." Around the time GPT-2 launched, we recognized that early language models could be weaponized for fraud. We turned the tables by using GPT systems to interact with scammers, extracting intelligence. Over time, these interactions became incredibly elaborate. For instance, our AI personas engaged scammers in year-long conversations sometimes lasting over 1000 messages back and forth, uncovering vast quantities of infrastructure being used for fraud, like crypto addresses and fraudulent bank accounts—in one conversation we uncovered 21 bank accounts being used for fraud from a single scammer. We had a policy that all outbound messages had to be human reviewed, which meant I had to manually read hundreds of thousands of LLM-generated and scammer messages, gaining a fair bit of insight into how basic LLM-powered systems behave: when they work, and when they don’t.
Pamir: There’s a lot of hype around AI agents. Can you explain where we stand today and where this technology is headed?
Joe: Defining AI agents is tricky, as the term is used broadly. I see agents as systems that can make decisions, interact with their environment, and execute actions autonomously. On one end, you have “traditional” LLM-based systems that make simple LLM calls to generate text. On the other, there are advanced agents capable of defining a plan, reasoning about their current world state, planning how to achieve their desired world state, adapting on the fly to novel situations and acting on the world using tools like code execution to achieve their objective —what I call "fully agentic systems”.
Today, most AI systems are closer to level 1 or 2 on a scale of autonomy. They perform specific tasks within strict constraints, like scheduling meetings or summarizing text. Moving forward, as agents become more sophisticated, they’ll tackle more complex, uncertain scenarios—like managing entire workflows or handling nuanced customer interactions. However, getting to that stage requires robust oversight and refinement.
Pamir: Is that where the Asteroid platform comes in? What problem does it solve?
Joe: Exactly. Asteroid is a platform for supervising and improving agents in real-time. Agents often face out-of-distribution scenarios–scenarios they haven’t seen before and have no guidance on how to solve. This is one of the reasons agents can be unreliable. At the end of the day, if you are an AI agent with code execution abilities and internet access then the world is your oyster and the consequences are unlimited! Asteroid allows these systems to learn from mistakes, with human oversight ensuring they adapt safely and effectively.
Our platform integrates various "supervisors"—humans, LLMs, or code—to catch and address failures. For instance, let’s say an agent in customer support encounters a complex query it mishandles. Asteroid flags the issue, escalates it to the right supervisor, and learns to avoid the same mistake. Over time, this process helps agents handle a broader range of scenarios independently.
Pamir: How do you see AI agents impacting industries? Will they replace jobs or augment them?
Joe: Agents will amplify human capabilities rather than replace them outright—at least initially. For example, instead of replacing customer support agents, they’ll help a single person overseeing multiple AI-driven interactions, stepping in only when necessary. This increases efficiency and allows humans to focus on more abstract, higher-value tasks. My belief is that it will be the job of humans over the next few years to move up the layers of abstraction and delegate more concretely defined tasks to agents which they oversee.
In the long term, as agents become more reliable, we may see companies run by a handful of people overseeing a network of autonomous systems. It’s a fascinating shift, but we’re still years away from that reality.
Pamir: How are you using AI to build your company in the backend?
Joe: I started using Cursor a few months ago and it has changed the way I approach prototyping and development, especially with unfamiliar languages or technologies. Before Cursor, I used Vim and VS Code, which were effective but limited. Cursor’s capabilities have made the initial stages of development remarkably efficient. During the early days of a project, I’d estimate my workflow is five to ten times faster. Even in later stages where more precision coding is needed, Cursor probably allows for about twice the efficiency of traditional IDEs, streamlining complex tasks and keeping momentum high.
I have also been experimenting with internal tools like Rebase, which started as a side project to solve a nagging problem: keeping our company wiki up-to-date. It pulls real-time updates from Slack, Jira, and other platforms, cutting out the manual upkeep. For me, this reflects the essence of AI-native company building: rapidly creating internal tools that address immediate needs while laying the groundwork for broader applications. Sometimes, these tools evolve beyond their initial scope and become standalone products with real market potential. It’s a blueprint for the future—iterating at speed, scaling when necessary, and staying laser-focused on utility.
Pamir: Shifting gears, what do you enjoy doing outside of work?
Joe: I’m a passionate climber and love being in the mountains. There’s something about being thousands of meters up, disconnected from technology, that’s deeply freeing. I love climbing in north Wales and in Scotland. One of my favorite climbs was Antisana in Ecuador, a ~6,000-meter volcano. Standing above the clouds, crossing crevasses, climbing ice, it was an experience that still inspires me.
Pamir: As a founder, how do you balance work and personal life?
Joe: Balance is crucial. I aim to be a founder who values both professional success and personal fulfillment. It’s easy to tie your identity entirely to your startup, but that’s a risky path—both mentally and emotionally. Having passions outside work keeps me grounded.
Pamir: Lastly, what trends or innovations excite you most beyond Asteroid?
Joe: Neurotechnology is an area I think is important. Companies are exploring brain-computer interfaces to enhance human intelligence and interaction with AI. It seems like there are obviously profound implications for the future. These advances could help merge AI with human cognition, and amplify our ability to reason and solve problems.
If you are building the next 10-person $100M ARR company, we want to talk to you.
Get in touch here: hello@blackbird.associates