Debate wins the agent bake-off, then loses to one model on the same budget

The load-bearing detail is which baseline you are shown. Debate won a contest among ways of arranging several agents: the study set it against sequential chains, ensemble methods and solo agents on multi-step reasoning ^[2], and among the team configurations debate finished first ^[1]. The same researchers then held total compute constant and found that one agent, working alone, frequently matched or beat the teams ^[3]. Both results sit in the same study. Only the first one sells a framework.

Stanford attributes the reversal to information lost when work passes between agents ^[4]. Treated as accounting rather than architecture, that means a share of every orchestration bill buys transmission instead of reasoning: tokens spent restating the problem to the next agent are tokens the single-agent configuration spent on the problem itself ^[12].

The scope of debate's win is narrow, and the report says so. Its clearest advantages came with less powerful underlying models, noisy or degraded input, and tasks requiring synthesis across large volumes of information ^[5]. The models under test were Qwen3-30B-A3B and Gemini 2.5 Flash, described as capable but below frontier class ^[6]. On a frontier-class model with clean, well-structured data, the report's own recommendation is a single agent, at lower latency and lower compute cost ^[10]. That makes debate a compensator for a weakness elsewhere in the stack, which is a reasonable thing to buy, and a different thing from a general capability gain.

The figure from the same report that will travel furthest is 37,000: a virtual laboratory of that many agents produced an antibody-drug conjugate design that Merck independently validated ^[9]. The published account gives no compute cost for that run and no single-agent comparison ^[14], so it evidences that agents at scale can reach a validated design, not what the design cost per unit of correctness against any alternative.

The report also carries the line that a mediocre model inside a well-designed debate framework can outperform a better model working alone ^[11]. That coexists with the equal-compute result only if the comparison ignores budget ^[13], since at matched budget the study puts the lone agent level or ahead ^[3]. Which yields a purchasing test the category does not advertise: take the total tokens an orchestrated run consumes, spend them on one model, on your own data, and compare. A framework that cannot clear that line is selling handoffs.