CMU Professor open-sources Agent framework Motus, with multi-model orchestration SWE-bench reaching 79% and halving costs

ME News report, April 15 (UTC+8). According to Beating monitoring, Lithos AI, an AI infrastructure company founded by Carnegie Mellon University Computer Science Department professor Dimitrios Skarlatos (CEO) and Zhihao Jia (CTO), has open-sourced its Agent service framework Motus under the Apache 2.0 license. The team is made up of researchers from CMU and Stanford, with members who have production infrastructure experience from AWS, Google, Meta, and NVIDIA.

Motus’s core idea is that different tasks suit different models. Rather than using the most expensive cutting-edge model to run every step, the system learns from the trajectories of production runs to automatically route different sub-tasks to the most suitable models. Currently, after Agent deployment, it remains static: the prompt framework, models, and context strategy do not change. Motus, by contrast, extracts task success-rate, latency, and cost signals from each run and continuously optimizes.

According to data from Lithos AI’s official website, on SWE-bench Verified, Motus’s multi-model orchestration achieves 79% accuracy—higher than Claude Opus 4.6’s 75.8% and GPT-5.3-Codex’s 72.6%—with costs less than half of using Opus alone. On Terminal-Bench 2.0, accuracy rises from 64% with Opus to 80.1%, and costs are reduced by roughly half as well. The framework also adjusts context memory strategies according to the specific workload and automatically detects steps that can be executed in parallel to reduce latency.

Motus is not tied to any particular model provider. It supports Agents built with OpenAI Agents SDK, Anthropic SDK, Google ADK, and pure Python, and provides Claude Code, Codex, and Cursor plugins. You can deploy locally with a single command or push it to the cloud. Free computing power is provided during the early preview phase.

(Source: BlockBeats)