Bardiya Akhbari Proposes Standard Vocabulary for Multi-Agent Research Systems

Bardiya Akhbari's new paper, 'A Vocabulary for Multi-Agent Automated Research Systems,' proposes a standardized language to describe and compare the architecture of autonomous research agents. The vocabulary defines eight core components: who the agents are, what operations exist, who can invoke them, how they communicate, what information is visible across runs, how the next action is chosen, how a run begins, and how outputs are evaluated. A trajectory captures one complete run from task input to final artifact. Because agents, operations, and initialization can be stochastic, repeated runs on the same task produce a distribution over trajectories rather than a single deterministic outcome. This formalization transforms vague architectural debates—such as when agents should communicate or carry state between runs—into empirically testable choices.

The paper also introduces two key measures of 'taste' in research systems. Generative taste quantifies the rate at which a system proposes novel trajectories before any feedback, while evaluative taste measures the gap between a proxy score and true quality. This separation clarifies common criticisms: poor generative taste means the system rarely explores unorthodox paths, while poor evaluative taste means its scoring mechanisms fail to capture real value. Akhbari demonstrates the vocabulary by instantiating it on several recent autoresearch systems, showing it accommodates widely differing designs. The work is relevant to AI, machine learning, and multiagent systems, and could become a foundational reference for building and benchmarking automated research pipelines.