One Researcher, Hundreds of AI Agents. Mirendil Raises $1 Billion to Build AI That Builds AI

What If AI Could Help Build the Next Generation of AI?

AI is already helping people write code, analyse information and conduct research. But a new category of companies wants to take that idea further: using AI agents to help develop and improve AI systems themselves.

San Francisco-based startup Mirendil has raised $1 billion in Series A funding to pursue that vision.

Announced on November 12, 2026, the round was led by Kleiner Perkins and Clearlake Capital, with participation from Battery Ventures, Doerr Capital, Gil & Co, Google, M13, Pear VC and S32, among other new and existing investors.

The Wall Street Journal reported that the round values Mirendil at approximately $5 billion. The company has not disclosed its valuation in its official announcement.

Founded by former researchers from frontier AI organisations, Mirendil is building tools designed to make advanced AI research and development more accessible to individual researchers, engineers and scientific teams.

Its central idea is straightforward but ambitious: AI agents should be able to take on more of the hands-on work involved in developing AI models, running experiments and improving systems.

The Experiment Behind Mirendil's Ambition

Mirendil's funding announcement describes an experiment that illustrates what the company is trying to achieve.

It says one researcher, working with AI agents, pretrained Mirendil-Base, a competitive coding model, from scratch in 3.5 months using less than $10 million in computing resources.

Traditionally, developing a model at this scale can require sizeable teams of researchers, engineers and infrastructure specialists.

Mirendil's approach is to use agents to handle much of the engineering and experimentation work, allowing a smaller research team to achieve more.

This is a company-reported result, rather than an independently verified benchmark. But it provides a concrete example of the kind of work the company wants its platform to support.

The broader ambition is to reduce the time, computing resources and human effort required to develop advanced AI systems.

Three Products Behind the Vision

Mirendil is building an ecosystem of models and agent-based tools rather than relying on a single AI product.

Its current offering includes three components:

Mirendil-Base: A coding model pretrained from scratch in the company's reported research experiment.

Fern: A model developed through post-training Mirendil-Base. The company says Fern supports context windows of up to 10 million tokens, allowing it to work with very large amounts of information across extended tasks.

Lor: An AI agent platform currently in private beta. Mirendil says Lor can coordinate hundreds of agents that write code, run experiments and iterate on results for days at a time. The company also uses Lor internally for its own research and development.

Together, these products reflect a broader shift from AI assistants that respond to individual instructions towards systems that can carry out longer, more complex workflows.

Who Is Building Mirendil?

Mirendil was founded by Behnam Neyshabur, its CEO, and Harsh Mehta, its CTO, alongside co-founders Shayan Salehian and Tara Rezaei.

The founding team includes researchers and engineers with experience at Anthropic, Google DeepMind, xAI and OpenAI.

Neyshabur previously worked at Google DeepMind, where he co-led the Blueshift team focused on reasoning capabilities in foundation models. He later joined Anthropic, where he co-led work on AI systems capable of long-term reasoning and research tasks.

Mehta also worked at Google and Anthropic, including on pre-training and automated AI research and development.

Mirendil's initial $200 million seed round was led by Andreessen Horowitz and Kleiner Perkins, with NVIDIA among the investors that subsequently backed the company.

The new $1 billion round significantly expands the capital available to the team as it develops its research infrastructure and agent platform.

Why Investors Are Betting on AI Research Automation

Much of the investment in AI has focused on building larger models, acquiring computing capacity and improving user-facing applications.

Mirendil is targeting another part of the value chain: the research and development process itself.

If AI agents can reliably write code, design experiments, analyse outcomes and help researchers iterate on models, they could potentially reduce the resources needed to develop new AI capabilities.

That could make advanced AI development more accessible to smaller teams and specialised research organisations that cannot afford the infrastructure and staffing available to the largest AI labs.

However, the technology still faces important challenges. Research requires more than running experiments quickly. Results need to be reproducible, reliable and independently validated. Human oversight will also remain important when agents are making complex technical decisions.

A $1 billion funding round gives Mirendil substantial resources to pursue its vision, but it does not establish that automated research can already replace large research teams.

The Bigger Question: Can AI Accelerate Its Own Development?

Mirendil is making a bet on a potential feedback loop: AI systems helping researchers build better AI systems, which can then help accelerate further research.

If that loop becomes reliable, it could change how quickly progress happens across software, engineering and scientific research.

But turning this idea into a widely used research platform will require more than impressive demonstrations. Mirendil will need to show that its tools deliver consistent results, work beyond its own internal environment and create measurable value for external researchers.

The bottom line: Mirendil is raising capital to tackle one of the biggest questions in AI: can AI meaningfully accelerate the process of developing more capable AI? Its reported experiment with one researcher and hundreds of agents points towards what may be possible. The next test is whether that approach can become a dependable tool for the wider research community.