Paper proposes an eight-part vocabulary for multi-agent research systems
Bardiya Akhbari · hf · 2026-07-28
A vocabulary for describing multi-agent research systems
This paper proposes a formal vocabulary for automated research systems made of one or more agents, so their designs can be described and compared more precisely.
It defines eight dimensions: who the agents are, what operations they can perform, who may invoke them, how they communicate, what state is visible across runs, how the next action is chosen, how a run starts, and how outputs are evaluated. A single run is represented as a trajectory, from task input to returned artifact.
The authors argue this framing makes previously fuzzy design questions testable, such as when agents should share information or gain capabilities. It also treats the evaluator as part of the system, because reported gains depend on how closely the proxy score matches real quality.
They further split “taste” into two different failures:
- Generative taste: how often a system proposes novel trajectories before seeing a score
- Evaluative taste: the gap between proxy score and true quality
They instantiate the vocabulary on recent auto-research systems to show it can cover designs with very different structures.
More from Research
- A ripgrep crash investigation ends up finding a kernel TLB-shootdown race — BenBajarin · 2026-07-28
- Curriculum learning can keep reinforcement-learning agents from exploiting random luck — ShawnHymel · 2026-07-28
- OpenAI and Anthropic may soon compete against their customers’ own data loops — bigdata · 2026-07-28
- SWE-rebench expands to five languages and adds local Qwen models — Fabulous_Pollution10 · 2026-07-28
- University of Basel is hiring PhD and postdoc researchers for LLM training — PMinervini · 2026-07-28
- Rednet could reuse much of Bittensor’s stack but ditch global consensus — sull · 2026-07-28