[Disclosure: AgFunder is the parent company of AgFunderNews.]
When we started AgFunder in 2013, we went out to raise a venture fund but didn’t have much luck as a first-time fund in a nascent sector. So we reverse-engineered the problem by building the infrastructure and ecosystem that would help us discover capital that was aligned with the future we saw in food and ag.
That infrastructure ended up becoming four pillars: content, code, community, and capital, which ultimately came last.
Our first thesis came from recognizing that the mainstream didn’t care about innovation in food and ag. Most VCs and founders were all chasing the next Facebook, Uber, or Airbnb, but we thought that there were others out there who felt venture capital could do something more substantial in the world; we just needed to find them. And so we created AgFunderNews as the TechCrunch for food and ag to show the world how things could be different.
Today, AgFunderNews reaches millions of readers a year, and about 95% of our LPs have come to us inbound. Trust can’t be rushed, and people often followed us for years before reaching out and writing a check. That was the first step.
Venture is an information business with amnesia
The second thesis we had was that if every company was going to become a tech company, then venture capital would be no different. So in 2015, we hired our first engineer to start building the foundation for this.
Over a decade, that system–which we call GAIA–has grown to track about 50 million companies; it sourced about two-thirds of our seed and pre-seed deals and today helps us diligence all our Series A and growth investments. It helps us identify potential investors when our portfolio companies go out to raise another round; it helps us with our due diligence; and it helps us better serve our LPs as they look for partners, market insights, and potential acquisitions.
Now we’ve entered the age of AI, and the only thing that can keep up with AI is AI, and we are executing on this as our third thesis by building out our internal agentic AI system, GAIA Brain.
For context as to why a VC partner is spending most of his time building AI at AgFunder today, I did my Master’s in AI in the early 2000s and then did my PhD at Yale in evolutionary computational biology, believing that a better understanding of biology and evolution would help break us out of AI’s long dark winter.
In 2007, I tried to convince my department to buy NVIDIA graphics cards to train neural networks. The answer was: “We’re not buying you gaming hardware.” In 2011, I applied for postdocs with Geoff Hinton (Just won the Nobel Prize for AI), Yann LeCun (later head of AI at Meta in 2012), and Andrew Ng (Became head of Google Brain).
All three wrote back some version of “No money, go do something else.” A year later, AlexNet was published out of Geoff Hinton’s lab, which bent the arc of history and brought us where we are today. I’ve been waiting for this moment a long, long time, and now it’s here.
The agentic inflection and my Skynet moment
Through the early AI and LLM years, our internal platform, GAIA, was largely a static knowledge base. Useful, but passive: you asked it a question, and it retrieved an answer. But it wasn’t a teammate, and it didn’t work toward goals.
Everything changed in early December 2025 with new agentic models and the release of OpenClaw. I mark that as the date we cracked open the door to the singularity. Suddenly AI could do things I only dreamed of doing six months ago.
To see how far we could push agentic workflows, I built type0.ai, a multi-agent newsroom designed to cover deeptech modeled off of AgFunderNews but with a broader mandate. Autonomous agents sourced stories, researched, drafted, fact-checked, and edited each other, running 60 to 100 model calls per article.
Today, type0 takes in over 5,000 signals per hour and can generate 1,000+ stories/day, burning 4-5 billion tokens. For reference, the Wall Street Journal and NYT produce about 500 articles/day, and a human knowledge worker uses about 10m-100m token-equivalent tokens PER YEAR. All this running on a single Mac Studio in my office.
Then came my Skynet moment. I gave the reporter agents a tool to request comments from sources so we could get original reporting and quotes for the articles, and tested it out by suggesting that the reporter (Sky) interview me (Rob Leclerc) for a story. It did so first by email, and then thirty minutes later, it had signed itself up for a voice-calling service and called me. By morning, the agents had contacted around a hundred real people, and real replies were sitting in the inbox.
I unplugged the machine.
That experiment proved two things:
- Agentic software is no longer passive; autonomy without a trust layer is unpriced risk.
- When research and drafting become cheap, the scarce resource becomes knowing what is credible, what changed, and what deserves human attention.
Enter GAIA Brain
Type0 was the stress test; GAIA Brain is the institutional memory.
GAIA Brain functions as a member of our team. It has read everything AgFunder has ever touched, every article, founder call, deal memo, and market map across a decade of work. Ask what a founder promised and needed two quarters ago, and it pulls up the exact conversation with source citations attached.
Ask who is building in a market, and it generates a map backed by years of continuous tracking. Ask which companies could be strong potential customers for a portfolio company, and it matches against Gaia’s data and our own data, then surfaces where we already have a warm relationship, through whom, and the context needed to activate it–all conducted in a simple thread, with no separate logins, passwords, or systems to navigate.
Parts of this already work: meeting prep, retrieving prior conversations, summarizing calls, tracking follow-ups, first-pass market maps, catching links between companies and people that would otherwise slip by. And parts of it still fail as we test it against real investment and partner workflows: incomplete retrieval, wrong tools, quitting too early, correct answers missing the context a person needs to trust them, and so on. Fixing those failures is most of the work right now.
But that work is also where the advantage is. Models are commodities; anyone can access an API. What is hard to copy is a decade-plus of proprietary history, domain depth, relationships, and the feedback that only comes from running agents inside real investment workflows. Every failure we fix is a lesson a general model never gets.
The future of venture
Near term, GAIA Brain helps AgFunder see more, forget less, and make better-supported decisions. Better meeting prep, better follow-through, faster research.
Further out, supervised agents maintain living views of markets, companies, and relationships. The market map stays current instead of getting rebuilt. And what our LPs get from us stops arriving four times a year and becomes a permissioned view that updates continuously.
AI is making action cheap, and cheap action will produce extraordinary companies and a flood of noise. The winners will be the firms that connect capable agents to proprietary memory without losing the source trail, the permission boundary, or the human point of view.
Venture still runs on trust, taste, ambition, and contact with the physical world, and AI raises the price of that judgment. That is what we are building at AgFunder: a firm that stops reconstructing context, starts exercising it, and gets a little better every day at knowing what deserves our attention.
This is the first in a series we’re publishing to document our progress as we build GAIA Brain in public: what’s working, what’s breaking, and what we’re learning along the way.
