The Defense Advanced Research Projects Agency (DARPA) is going all-in on AI agents that don’t just trade answers, but could help unlock the next wave of scientific breakthroughs.
Its new Mathematics of Boosting Agentic Communication (MATHBAC) program sets aside up to $2 million in Phase I funding under a 34-month effort to rethink how AI systems collaborate from the ground up.
It targets AI’s reliance on trial-and-error — what the agency calls an “Edisonian” approach — where systems brute-force solutions without fully understanding why they work.
That limitation shows up in multi-agent setups, where AI can coordinate and produce results, but still lacks a clear grasp of the reasoning behind them.
“While AI excels at navigating solution spaces, it struggles to systematically explore hypothesis spaces,” DARPA said.
The MATHBAC Challenge
DARPA splits the problem into two under MATHBAC’s 16-month Phase I.
First, researchers will work on the math behind communication protocols of AI agents — who talks, when, and how.
Second, they will focus on what gets shared, aiming to extract “compact, generalizable ‘nuggets’” of knowledge from raw data.
Beyond that, DARPA wants evidence that agents can derive principles from data that predates their original discovery, recreating concepts as foundational as the periodic table purely from data, and potentially going further.
Phase II raises the stakes: agents will be expected to evolve their own communication strategies, knowledge bases, and even emergent languages.
During this period, DARPA envisions a shift in “evolutionary pressure” from human developers and toward the AI systems themselves.
Throughout the program, the agency will prohibit small tweaks and other incremental upgrades to force more fundamental advances.
Proposals are open to industry partners until June, with work expected to begin in September.
