The article evaluates a routing model from 基元律动, referred to in the title as OpenSquilla, against Jev on 585 historical AppWorld task replays. The router selects among four execution tiers, C0 through C3, before another model carries out tasks such as email, shopping, payment, and cross-application operations. It completed 522 tasks, compared with 519 for Jev, raising the completion rate from 88.72% to 89.23%, or 0.51 percentage points. Average downstream generation cost fell from $0.181259 to $0.085974 per task, a 52.57% reduction; the calculation excludes routing, evaluation, and deployment costs. Local CPU routing took an average of 45.09 ms, with a P95 latency of 50.10 ms, while Jev’s remote API call averaged 820.81 ms. The routing classifier has about 0.95 million parameters and can run locally on a CPU. Its training data was built from simulated agent executions across different scenarios and difficulty levels; three models independently labeled uncertain samples, with the median label used and more consistent labels receiving greater weight. Against always using Kimi-K3, the router improved task completion by 4.19% relatively and reduced generation cost by 60.83%. Three examples illustrate the trade-off: it selected cheaper or lower-tier models that completed all required actions, while Jev or its selected model missed playlist items, omitted an email recipient, or stopped after writing a verification script without executing it. The examples were drawn from 18 tasks won by the router and 15 won by Jev, leaving a net advantage of three tasks, so the overall completion difference remained modest despite the larger cost and latency gains.
