Jev as a dispatcher for multiple agents: fewer tokens, similar accuracy

Four researchers from Shanghai Jiao Tong University published a preprint called Jev-LiteMAS on 15 October 2026. S56 It asks whether, in a system of several AI agents, an expensive generative model has to decide every step, or whether a decision model can handle the coordination. The result is relevant to anyone who builds agents and pays for tokens: in the authors' test, Jev took over routing and the first check of intermediate results, generative-model token use fell, and accuracy stayed close to the comparison system. This is a preprint that has not yet been peer reviewed.

Where agents spend their tokens

The authors first measured the existing multi-agent frameworks MetaGPT, AgentNet and LAS with the DeepSeek-V4-Flash model. Across six measured configurations, routing and review accounted for 31.7 to 43.5% of generative-model tokens and 29.5 to 46.1% of the time. S56 So this is not the work itself but the overhead around it: who goes next and whether an intermediate result is good enough.

They also acknowledge that a decision model does not replace a generative one one-for-one. In typical systems, a planner writes a multi-step plan and instructions for the next agent, while a reviewer explains what needs fixing. Jev can do neither; it only returns a selection from options or a verdict. S56

How the work is split

The design therefore divides tasks by the kind of output they need. After each step, Jev picks which of the permitted experts continues; if the system allows termination, STOP is among the options. Nobody writes a full plan in advance. At designated points, Jev assesses an intermediate result and decides only "pass" or "escalate". The generative model does the actual work, and only what Jev escalated gets a second review; its feedback then goes to the expert who fixes the error. S56

The system is built on the AutoGen framework, with DeepSeek-V4-Flash and GPT-5 as the generative models. The authors evaluated it on the MATH, MBPP+, API-Bank and IDRBench I3 tasks, using 150 samples for each (for API-Bank, 150 episodes). The preprint does not state which Jev version was used. S56

What the numbers showed

Compared with LAS, which the authors use as the main comparison, generative-model token use fell by 12.7 to 54.6%. Accuracy differed by no more than 1.34 percentage points, and latency was lower in seven of eight settings. The exception was the MATH task with GPT-5, where the system needed 30.04 s instead of 25.91 s. S56

The largest difference came on the IDRBench I3 task with DeepSeek: 10,233.8 instead of 22,556 tokens and 72.40 instead of 133.52 seconds, at an accuracy of 75.33% versus 76.00%. S56 These counts do not include Jev's own tokens; they cover only the generative model.

The comparison of routers on the same task is instructive. The Jev router reached 75.33%, a generative-model router 68.67%, a small language model 64.00% and fixed rules 26.67%. A variant that did not see the experts' outputs when routing ended at 57.33%. S56 Decision-making therefore helps most when the model sees what the previous step actually produced.

In review, Jev was slightly less accurate than the generative reviewer. On the MATH task it caught 95.33% of faulty intermediate results (generative model 96.67%) and needlessly escalated 4.00% of correct ones (generative model 1.33%). On MBPP+ the figures were 91.33% versus 92.00% and 6.67% versus 5.33%. S56 With GPT-5, Jev itself added 1.13 to 2.21 s per task, or 2.92 to 6.85% of total time, and made up 1.56 to 2.14% of model costs. S56

Limits the authors acknowledge

The experts are fixed in advance, and the system does not create new roles. The savings from two-stage review shrink when more outputs need escalation or repair. The authors openly describe the main reliability risk: an error that Jev lets through is never seen by the generative model. S56 The results also rest on 150 samples per task, which is a modest sample for differences of around one percentage point.

What a Slovak project can take from this

If your Slovak online shop has automation that chains several steps, for example recognising the type of a return claim, looking up the order, drafting a reply and checking it, the preprint shows a sensible division of labour. Deciding "which step comes next" or "is this draft acceptable" is a Choice question, while writing text stays with the generative model. TypeSafe charges 0.042 USD per million input tokens, output is free, and the stable alias jev-latest points to jev-1.13.0. S01 The cost calculator shows what that would cost at your volume.

The review figures also carry a warning. In this test, Jev let slightly more errors through than the generative reviewer, and nobody checked those afterwards. For steps with financial or legal consequences, such as refunds or cancellations, we therefore think the final say belongs to code or a human. Check on your own data when to escalate; the guide on confidence and calibration can help. An earlier study of Jev in agents offered a similar lesson: a faster decision does not by itself automatically lead to a better result.