DARPA Is Betting $2 Million That the Agent Stack Needs a Theoretical Foundation
The protocols exist. MCP, ANP, and a handful of competing specifications are racing to define how AI agents discover each other, route requests, and share context across a network. These are real problems with real engineering solutions. But DARPA, the Pentagon's research arm, is asking a different question: what if the plumbing is being built before anyone understands the physics?
MATHBAC, the Mathematics of Boosting Agentic Communication, launched April 7, 2026 with up to $2 million per Phase I award and a June 16 proposal deadline. The program runs 34 months split across two phases, and its ambition is to do for agent-to-agent communication what information theory did for data transmission: provide a mathematical substrate that lets you reason about correctness, efficiency, and generalization, rather than just hacking until it works.
The first technical area is approachable. MATHBAC wants to develop the mathematics behind agent communication protocols — formal methods for verifying that a multi-agent system does what you intended, not just what it happened to do in testing. The second is stranger. DARPA wants performers to extract generalizable principles from what agents actually communicate: laws, correlations, compact representations of knowledge that become part of a shared reasoning substrate rather than evaporating when the session ends.
DARPA's illustration of this goal is a red flag, not a promise. The program asks whether a collective of domain-specific agents could, from raw data, collectively rediscover the periodic table and then extend it into a multidimensional analog for molecules. The Register noted DARPA itself describes this as "generally considered nigh impossible." Nobody expects a working result in 34 months. The example is useful precisely because it fails — it exposes what current agent frameworks cannot do: systematic hypothesis generation where principles discovered by one agent become part of a shared reasoning substrate that others build on.
The distinction this reveals matters for anyone building in the space. Most agent infrastructure work right now is concerned with transport: how agents exchange messages. MATHBAC is trying to solve the semantic layer. What should agents be saying to each other? How does shared understanding emerge from collective reasoning rather than human programming? Right now, those questions are answered by whoever writes the most convincing API spec.
DARPA frames the gap without diplomatic hedging. Current AI development, including agent frameworks, remains heuristically guided, an ad hoc trial-and-error process focused on outcomes rather than understanding why those outcomes occur. Agent-to-agent communication, without a rigorous mathematical foundation, will remain inefficient, inconsistent, and difficult to generalize across domains.
Phase I runs 16 months and covers protocol mathematics and content extraction. Phase II, 18 months, asks something harder: can agents evolve their own cooperation skills? DARPA wants to shift evolutionary pressure from human developers onto the agents themselves. Performers in Phase II will work in what the solicitation calls Evolution Teams: collectives of agents optimized not just to solve problems but to improve how they solve problems together. This is where the program brushes up against something genuinely open in AI research — whether a population of agents can compound reasoning over time the way scientific communities do, or whether each deployment is essentially a reset.
MCP and ANP are building the pipes. They solve real problems: how you connect a dozen tools to an agent without custom code, how you route requests across a network. MATHBAC does not compete with these protocols. It asks whether the agent economy, once it reaches a certain scale and complexity, will need a theoretical foundation that protocol engineering alone cannot provide. The answer depends on whether you think agent collectives are a scaling story or a reasoning story. If the former, the current stack is probably fine. If the latter, you need the math.
The program starts September 2026. Proposals are due June 16 at 4:00 PM ET. DARPA has said explicitly that incremental improvements to existing approaches will not be competitive — it wants novel foundational work in mathematics, systems theory, and information theory applied to agent communication and collective scientific reasoning. Whether the bet pays off is genuinely unknown. But the framing deserves attention: building the transport layer before you understand the semantics is a reliable way to end up with very fast roads to nowhere.
Sources: The Register | DARPA MATHBAC program page | BW&Co consulting summary
