An Oct. 1 arXiv preprint from researchers at Oxford, KAUST, Sydney and Cambridge tests a different way to coordinate AI scientists: the active agent chooses which existing specialist works next, based on the project's current state. That setup had the highest average benchmark score in each of three tested systems. But the evaluation used one seed, so it does not show whether the pattern would hold across independent runs.
AI research systems often split literature review, experiments and writing among specialized agents. A fixed workflow decides who acts at each stage. The paper asks whether a system can choose its next worker after seeing what the current team has done and what remains unfinished.
The authors call their method Runtime Agent Coordination, or RAC. At each handoff, the active agent chooses another specialist already allowed by the system, using the project's saved files, unresolved problems, work history and remaining budget. RAC keeps each host's existing models, agents, tools and permissions.
The authors compared four versions of each host's workflow. The native version follows its usual process. A second lets agents communicate but keeps the same handoff order. A third lets the active agent choose who works next. A fourth also gives that worker a scoped contract and checks its saved output; the check becomes feedback for a later agent rather than a reason to block or undo work.
On ResearchClawBench, Agent Laboratory and EvoScientist were each tested on 10 tasks, while ARK was tested on five. Each task provides a research question, related papers, raw data and an executable environment, while withholding the target paper. The agents used DeepSeek V4 Pro, and an external GPT-5.5 judge scored their final research outputs.
Letting the active agent choose the next worker produced the highest mean within each system: ARK scored 18.42 versus 16.66 for its native workflow; Agent Laboratory scored 12.08 versus 5.47; and EvoScientist scored 18.53 versus 15.99. These are comparisons within each system, not a ranking across systems, which used different task sets.
Adding contracts and artifact checks together lowered those three means to 17.98, 9.88 and 15.96. Because the authors changed both features at once, the comparison cannot show which one explains the drop. More seeds, tasks and models are needed to test how stable the result is.
RAC changes who acts next inside a research project, not which research question the system pursues. The result makes the handoff decision itself a design choice worth testing across more tasks and independent runs.