Veris.ai – Why the Environment is Everything for Autonomous AI Agents

Today, we are announcing our investment in Veris.ai, founded by Andi Partovi and Mehdi Jamei. You might have heard that 2025 is the “year of the agents”, but no one seems to agree what an agent really is. The piece everyone does agree on is that an agent lives in some sort of an environment, and has to take actions on it either through tool calls or MCPs. Veris helps companies build simulation environments that they can use to train and test agents before they go to production, using a mix of synthetic scenarios, verifiable rewards, and production logs. After running the agent in the environment in thousands of simulations, they use reinforcement learning to improve your agent performance in the tasks that matter the most. We spoke with Andi and Mehdi about their vision for Veris in our founder Q&A:

What was your first experience with computers?

Andi: I grew up in Tehran, and my first exposure to computers came in pre-school through a neighbor who owned a PC. We'd play "Prince of Persia" and various board games together. One day, I started programming one of those board games on the computer—that sparked my lifelong passion for programming, particularly videogames.

Mehdi: I grew up in Neyshabur, and my first computer was a Commodore 64 I received in 4th grade. Armed with a BASIC programming book, I started teaching myself coding. Growing up in a lower-middle-class family, upgrading to a Pentium later on was a significant step up. I eventually taught myself C++ and Pascal, further deepening my love for coding.

When did you first get entrepreneurial experience?

Andi: As a kid, I was deeply into puppet shows and had a big collection of puppets. My cousin handled the business side—making and selling tickets—while I acted out stories. It was my first experience blending creativity with a business mindset.

Mehdi: In the summer after first grade of high school, a friend and I began selling English-to-Farsi translation software. We'd burn open-source software onto CDs and sell them door-to-door. The money we made funded a memorable trip, marking my earliest entrepreneurial venture.

When was your first "wow" moment with LLMs?

Mehdi: At System, we were building numerous NLP models. When GPT-2 came out, it was impressive but didn't quite reach the "wow" threshold. However, GPT-3 was mind-blowing; it enabled us to replace multiple models at once, drastically simplifying our workflow.

Andi: Our lab had been developing many statistical language models—mostly non-generative. The introduction of transformer-based models were a clear game changer, fundamentally shifting how we approached model building.

Why did you decide to start Veris.ai?

Andi: Before Veris, I was working on an autonomous marketing agent. Generation quality was good, but chaining multiple generative steps together consistently was incredibly difficult. I saw a clear need to build the necessary infrastructure to support more sophisticated agent capabilities.

Mehdi: We explored various application-layer ideas but quickly realized the major bottleneck wasn't applications—it was infrastructure. Autonomous AI agents require realistic environments to evaluate, train, and test effectively. Today, most companies jump straight from model-building to production without the necessary intermediate environments. We created Veris.ai specifically to solve this infrastructure gap.

Will 2025 be the year of autonomous AI agents, or is more time needed?

Mehdi: We're optimistic, but realistic. At Workmate, we tried hard to minimize the human-in-the-loop component, aiming for 30-40%. But the reality is that many teams still rely on humans for 80-90% of their processes. Bridging this gap won't be instant. Reinforcement learning (RL) is essential to achieving near-perfect accuracy, but it's still out of reach for most teams. While hitting 80% accuracy might be straightforward, reaching 90% is challenging, and 99% accuracy is practically impossible without significant advances in RL.

Andi: Beyond technical barriers, there's also a psychological hurdle. Companies accustomed to the SaaS mindset often find it difficult to fully embrace autonomous agents. To truly leverage AI, organizations need to stop building mere copilots and fully commit to the autonomous agent approach. This cultural shift will take time.

What is your hot take on building AI agents?

Andi: My controversial opinion is that there shouldn't be a separation between training, staging, and production environments for AI agents. Ideally, these should be one continuous environment. Begin with 100% synthetic data and environment feedback, then progressively introduce real-world data until the environment closely matches production reality.

Mehdi: Most teams approaching AI agents come from a traditional software engineering background. However, building autonomous agents demands a radically different mindset—one focused on creating realistic, responsive environments rather than writing static software. Ignoring this critical distinction is where many teams fail.

We're excited about the future Veris.ai is building and can't wait to see how Mehdi and Andi continue to shape the world of autonomous AI agents.