DARPA launches MATHBAC program for multi agent AI research

A top-secret US government body called the Defense Advanced Research Projects Agency (DARPA) published a formal program announcement on April 7, 2026, launching a new research initiative called the Mathematics of Boosting Agentic Communication, or MATHBAC, which aims to develop the mathematical and scientific foundations needed to make networks of artificial intelligence agents collaborate more effectively — and ultimately accelerate the pace of scientific discovery for national defense.

The solicitation, issued by DARPA’s Defense Sciences Office seeks proposals from universities, research institutions, and private companies willing to tackle one of the deeper unsolved problems in modern AI: not how to make a single AI system smarter, but how to make groups of AI agents communicate, coordinate, and collectively arrive at scientific insights that no single system — or even a single human expert — could reach alone. Proposals are due June 16, 2026, with program performance expected to begin September 15, 2026.

The program is structured as a 34-month, two-phase effort. Phase I, running approximately 16 months, will cap individual awards at $2 million and focus on developing the mathematical tools to analyze and improve how AI agents exchange information. Phase II, spanning the following 18 months, will push further — asking whether AI agents can be designed to evolve their own communication strategies and knowledge bases as they work, moving toward what DARPA describes as autonomous generation of new science itself.

MATHBAC is a response to a recognized gap in how AI systems currently operate. Today’s most capable AI platforms — large language models, reasoning models, and specialized science models — tend to be designed and optimized through trial-and-error processes that DARPA’s program documents describe as fundamentally “Edisonian”: testing combinations until something works, without a rigorous theoretical understanding of why. The agency wants to replace that approach with a principled mathematical framework, drawing on control theory, information theory, and applied mathematics to describe agent behavior as precisely as engineers describe physical systems.

The program is divided into two technical areas. The first, called TA1, focuses on the mathematics of communication protocols — essentially, the rules that govern which AI agent talks to which, when, and what it says. Researchers working in this area would develop tools to model individual AI agents as mathematical operators, analyze how different communication structures affect a team’s ability to solve problems, and build software that lets researchers design optimized protocols without large-scale trial and error. The second area, TA2, focuses on what agents actually share with each other: specifically, whether AI systems can extract and articulate genuinely new scientific principles from large volumes of data, the way scientists once derived fundamental laws from experimental observation. A stated goal for TA2 — described in the solicitation as a “DARPA-hard” challenge — is to demonstrate that an AI agent collective could, in principle, rediscover something as foundational as the periodic table purely from data, before that organizing principle had been explicitly named.

DARPA is restricting MATHBAC’s focus to scientific domains deliberately. Science fields offer structured, well-curated bodies of knowledge; they require the kind of interdisciplinary synthesis that is difficult for a single agent; and they provide clear standards for validating whether a claimed discovery is real. The solicitation mentions chemistry and biomedical research as illustrative examples, while leaving proposers free to choose their own science subdomains.

MATHBAC fits within a broader pattern of Pentagon interest in harnessing AI for scientific research at speed. The DARPA solicitation itself notes recent government initiatives — including the Department of Energy’s Genesis Mission, a collaboration with multiple AI companies — as evidence that autonomous AI-assisted discovery is already attracting significant institutional investment. Where those initiatives have focused on building and deploying agentic platforms, MATHBAC is explicitly aimed at understanding why such systems work, or fail, at a mathematical level — a distinction DARPA frames as essential if the field is to move beyond ad hoc successes toward reliable, generalizable capability.