darpa is funding the math of multi-agent communication — here's the governance gap that creates

darpa's MATHBAC program (Mathematics of Boosting Agentic Communication) opened proposals in april with up to $2 million in phase I funding over 34 months. proposals due june 16, program starts september 2026. the focus: foundational mathematics for agent communication protocols, multi-agent coordination science, self-evolving agent systems.

government-funded multi-agent research is interesting for a few reasons, but the one worth thinking through is the governance problem it creates downstream.

what MATHBAC is building toward

the program is explicitly about the mathematical foundations of how agents communicate — not specific implementations, but the protocol-level primitives that determine what agents can express to each other, how they coordinate on tasks, and how they evolve communication patterns autonomously.

that last part is the one that makes compliance teams nervous. "self-evolving agent systems" is a research framing for systems that update their own behavior based on experience. from a regulatory standpoint, that's a high-risk AI system under EU AI Act Article 6 criteria — it affects consequential decisions and its behavior is not fully specified in advance.

the audit problem with self-evolving systems

static agent systems are already hard to audit. you need to document what the model was, what its authorization scope was, what decisions it made, and trace those decisions to the output.

self-evolving systems make that harder by design. the system's effective behavior at time T+1 is a function of its experience at time T. if you can only document the initial configuration, you haven't documented what the system actually did by month three.

MATHBAC is funding research that will eventually produce production deployments. those deployments are going to need audit infrastructure that tracks not just individual agent decisions but the evolution of the agent's communication and coordination patterns over time.

where the market gap is

right now, audit tooling for AI agents treats agent behavior as static — you document the model, the prompt, the tool permissions, the outputs. that's the right starting point for current deployments. it's what BizSuite AI-Audit does: 48-hour delivery, $997 entry point, structured decision-trace documentation for EU AI Act compliance.

what MATHBAC research will eventually require is an audit layer that can track behavioral drift — the delta between how an agent was documented at deployment and how it's behaving now. that's a harder problem, and nobody's selling it yet.

the MATHBAC RFP is a leading indicator for where the compliance problem is going to be in 18-36 months. the teams building audit infrastructure now — for static deployments — are the ones who'll be positioned to extend it when self-evolving systems hit production.

the starting point for that infrastructure: https://getbizsuite.com/ai-audit

NOTE: DARPA is flagged as a government agency with a long sales cycle. this article is written as thought-leadership positioning, not direct outreach to DARPA. publisher to confirm the right distribution channel (Dev.to vs own blog) before ship.