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

ME News Report, April 15 (UTC+8), according to Beating Monitoring, Carnegie Mellon University Computer Science Professor Dimitrios Skarlatos (CEO) and Zhihao Jia (CTO) founded AI infrastructure company Lithos AI, which has open-sourced the Agent service framework Motus under the Apache 2.0 license. The team consists of researchers from CMU and Stanford, with members experienced in production infrastructure at AWS, Google, Meta, and NVIDIA.
The core idea of Motus: different tasks are suited to different models. Instead of always running all steps with the most expensive cutting-edge model, the system learns from production run trajectories to automatically route sub-tasks to the most appropriate models. Currently, after deployment, Agents are static, with fixed prompt frameworks, models, and context strategies. Motus extracts signals such as task success rate, latency, and cost from each run to continuously optimize.
According to Lithos AI’s official website data, 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 increased from 64% with Opus to 80.1%, with costs similarly halved. The framework also adjusts context memory strategies based on specific workloads and automatically detects steps that can be executed in parallel to reduce latency.
Motus is model provider-agnostic, supporting OpenAI Agents SDK, Anthropic SDK, Google SDK, and Agents built purely in Python. It offers Claude Code, Codex, and Cursor plugins, deployable locally with a single command or pushed to the cloud. Early preview phase provides free computing power.
(Source: BlockBeats)