From Researcher to Orchestrator: Attention-Bounded Supervision of AI Agent Teams in Engineering Research

Researchers in electrical and computer engineering increasingly delegate literature search,coding, data analysis and drafting to AI agents, yet practices for supervising several agents atonce lag behind adoption, and the researcher remains accountable for every claim. Existingmulti-agent frameworks and levels-of-automation models do not tell a researcher how muchautonomy to grant a task, how work should pass between agents, or how many agents one personcan supervise. This paper presents an orchestration framework that answers these questions.Each task’s autonomy level is set from three properties of its effects (reversibility, reach andverifiability) by a rule that is complete and monotone. Work passes between agents as version-controlled files that no agent may mark as accepted when it produced them, and every reportednumber must name the computation that produced it. A review-load model bounds team size bythe researcher’s daily attention and shows that the ratio of planning to execution agents followsfrom their review costs rather than being fixed. We apply the framework to an illustrative analog-to-digital converter characterization task and to two AI-assisted publication projects, a 20-chapter textbook and an arXiv paper, whose logs and repositories record 13 defects. Mechanicalchecks caught every defect for which one existed, errors of meaning were caught only whensomeone recomputed or read the work, and a quality score that no computation had producedescaped every control and was published. The results suggest measuring review load and escapeddefects before enlarging an agent team.