DARPA’s MATHBAC Project: Boosting Agentic AI Communication via Mathematical Code

The U.S. Defense Advanced Research Projects Agency (DARPA) is venturing into a novel frontier of artificial intelligence, seeking to move beyond simple prompts and responses toward a world where AI agents can communicate with one another using a sophisticated mathematical language. The initiative, known as the Mathematics of Boosting Agentic Communication (MATHBAC), aims to create a framework where agentic AI platforms can collaborate, share information, and fundamentally understand the mechanics of their own operations.

This effort represents a strategic shift in how the U.S. Government views the evolution of AI. Rather than focusing on incremental updates to existing Large Language Models (LLMs), DARPA is funding foundational research to derive the science of AI communication. By leveraging information theory and systems theory, the program intends to enable AI agents to disentangle complex data streams and extract generalizable principles, effectively transforming how scientific discovery is conducted for defense purposes.

The scale of the ambition is reflected in the funding and the strict requirements for a “breakthrough” approach. DARPA is offering Phase I awards of up to $2,000,000 per award for teams capable of developing new mathematical frameworks and algorithms for multi-agent AI systems. The agency has explicitly stated that research resulting in only incremental improvements to existing methods is excluded from this funding opportunity.

Decoding the MATHBAC Framework

At its core, MATHBAC is designed to solve a critical gap in current AI capabilities: the lack of a formal, mathematical governing system for how multiple AI agents interact. Whereas current “agentic” workflows often involve one AI calling another via a set of pre-defined instructions, DARPA is looking for a more organic, mathematically derived method of collaboration. This involves the use of mathematics, systems, and information theory to govern how agents collaborate and share information to extract generalizable principles.

The program is structured to move from theoretical derivation to the creation of practical tools. The first phase focuses on the derivation of the mathematics behind agentic AI and improving communication between systems. The second, more ambitious phase aims to create tools that enable the development of a new science, solving the fundamental mathematical problems that underpin collective agentic intelligence.

This approach is not merely about making AI “chat” more efficiently. This proves about enabling platforms to understand the fundamentals behind their own functioning and the solutions they produce. By formalizing this communication, DARPA hopes to accelerate the pace of scientific discovery, allowing AI agents to work together to solve complex problems that are currently beyond the reach of single-model systems.

Key Objectives of the Research

To achieve these goals, DARPA has outlined several specific areas where foundational research is required:

• Mathematical Frameworks: Developing the formal rules and protocols that govern how AI agents communicate.
• Systems and Information Theory: Advancing the theory of how AI agents collaborate and share information efficiently.
• Generalizable Principles: Creating algorithms and tools that allow agents to extract broad scientific principles from complex data streams.
• Simulation Tools: Building software tools to simulate and optimize how these agents interact before they are deployed in real-world scenarios.

The program is managed by the Defense Sciences Office (DSO), with Yannis Kevrekidis serving as the program manager for the initiative.

Timeline and Application Milestones

The MATHBAC program is operating on a tight schedule, with several critical deadlines for researchers and organizations looking to participate. The solicitation for the program was published on April 7, 2026 by DARPA.

For those interested in applying, the timeline is as follows:

The agency expects to achieve its primary goals within a 34-month window, signaling a rapid push toward a new era of collective agentic intelligence.

Why This Matters for the Future of AI

The implications of the MATHBAC project extend far beyond defense. If AI agents can be taught to communicate through a formalized mathematical code, it could lead to a paradigm shift in how we approach “black box” AI. Currently, the reasoning processes of deep learning models are often opaque. Although, if the communication between agents is governed by a transparent mathematical framework, it may become easier for humans to audit, verify, and understand how these systems arrive at specific scientific breakthroughs.

this research addresses the “silo” problem in AI. Currently, different AI models are often limited by their specific training data or architecture. A universal, mathematically grounded communication protocol would allow diverse agentic platforms to collaborate regardless of their underlying build, much like how the internet allowed disparate computer systems to share data decades ago via a common protocol.

By focusing on “basic and applied research” rather than “commercialization” or “product development,” DARPA is ensuring that the resulting breakthroughs are foundational. This ensures that the resulting science can be applied across various domains, from medicine to materials science, while maintaining a primary focus on national security and defense capabilities.

The next major checkpoint for the program is the proposal abstract deadline on April 30, 2026. We will continue to monitor the Defense Sciences Office for updates on the selected research teams and the subsequent rollout of Phase I results.

Do you think mathematical formalization is the key to solving the AI “black box” problem? Share your thoughts in the comments below.