CMU Professor Opensource Agent Framework Motus, Multi-Model Orchestration SWE-bench Reaches 79% and Halves Costs

ME News: On April 15 (UTC+8), according to Beating Monitoring, Motus, an open-source Agent service framework under the Apache 2.0 license, was released by Lithos AI—an AI infrastructure company founded by Dimitrios Skarlatos (CEO) and Zhihao Jia (CTO), professors in the Department of Computer Science at Carnegie Mellon University. The team is made up of researchers from CMU and Stanford, with members experienced in production infrastructure from AWS, Google, Meta, and NVIDIA.

The core idea behind Motus is that different tasks are suited to different models. Instead of always running every step with the most expensive cutting-edge model, the system learns from the trajectories of production runs and automatically routes different sub-tasks to the most appropriate models. At present, after deployment, the Agents are static: the prompt framework, models, and context strategy remain unchanged. Motus extracts signals—task success rate, latency, and cost—from each run to continuously optimize.

According to data from Lithos AI’s official website, on SWE-bench Verified, Motus’s multi-model orchestration reaches 79% accuracy—higher than Claude Opus 4.6’s 75.8% and GPT-5.3-Codex’s 72.6%—and the cost is less than half of using Opus alone. On Terminal-Bench 2.0, accuracy increases from Opus’s 64% to 80.1%, with costs similarly reduced by about half. 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 specific model provider. It supports OpenAI Agents SDK, Anthropic SDK, Google ADK, and Agents built purely with Python, and provides Claude Code, Codex, and Cursor plugins. It can be deployed locally with a single command or pushed to the cloud. During the early preview phase, computing power is provided for free.

(Source: BlockBeats)