Broader impact
Why this research matters beyond SwarmEngines.
Six dimensions of broader impact, each tied to a Phase I deliverable.
Sector reach
33.3 million U.S. small businesses.
The U.S. small-business sector accounts for 99.9% of all U.S. firms and employs roughly 46% of private-sector workers. It is the part of the economy with the least capacity to safely adopt autonomous AI on its own — no internal security teams, no compliance staff, no engineers to interpret an audit log.
U.S. economic competitiveness
Closing the productivity gap that mid-market AI cannot reach.
Autonomous agent platforms today are sold to enterprise customers at price points and with custom-installation profiles that exclude any business below ~250 employees. Without research on the unit economics and safety floor of multi-tenant agent fleets, the small-business sector is structurally locked out of the same productivity gains the F500 is realizing this decade.
AI-safety surface
A trustworthy-AI baseline that does not exist today.
If autonomous agents are deployed to the SMB sector with the architecture mainstream platforms ship today — shared runtime, shared credentials, no coordination supervision, opaque automation — the resulting failure surface will be measured in compromised customer-payment systems at HVAC firms, leaked health data at dental practices, and contract violations at small legal firms. The research described elsewhere on this site is the safety floor required for the deployment to happen responsibly.
Knowledge dissemination
Phase I deliverables go to peer-reviewed venues and will be open-sourced where commercially compatible.
We have no incentive to keep the safety floor proprietary. Our defensibility is in the integrated runtime + skill catalog + customer relationships — not in the safety supervisor or the audit-chain primitive. The runtime supervisor evaluation harness (Problem 2 deliverable) and the audit-chain reference implementation (Problem 3 deliverable) are particularly good candidates for open release because the entire industry benefits from a shared baseline.
Workforce effects
Job augmentation at the small-business owner-operator level.
Our agents are designed to run a single business owner from working 70 hours/week down to working 40 hours/week — by absorbing the call-back, follow-up, scheduling, review-reply, and admin layers that owner-operators currently personally do at 10pm. This is augmentation, not displacement: the businesses we serve already do not have the head-count to displace.
Educational impact
A teaching artifact for multi-agent safety.
Multi-agent LLM coordination is currently taught from idealized game-theoretic frameworks (Hadfield-Menell, Stuart Russell's coursework) without a production runtime artifact students can study. Our open-sourced Problem-2 evaluation harness (200 labeled multi-agent scenarios with conflict outcomes) can serve as a course resource for AI-safety curricula. We will deposit it with at least one academic collaborator as part of Phase I.
Continue reading