New Vocabulary for Multi-Agent Automated Research Systems

A newly established vocabulary has been created for the purpose of describing and assessing automated research systems composed of one or multiple agents. This vocabulary outlines eight essential design decisions: agent identity, permitted operations, invocation rights, communication techniques, visibility of information during and between runs, action selection, initialization of runs, and evaluation of outputs. Each trajectory captures a single run from the input task to the resulting artifact. Due to the stochastic nature of agents, operations, and initialization, multiple runs on the same task yield a distribution of trajectories. This vocabulary transforms structural design inquiries—like the timing of agent communication and the transfer of information—into verifiable choices, incorporating the evaluator as a system component, as reported outcomes are influenced by the evaluator's alignment with user objectives.