Silicon Valley Chinese Team AFK AI Open-Sources Super-agent Architecture Comma for 24/7 Parallel Task Processing

Silicon Valley-based Chinese team AFK AI has announced the open-source release of its Super-agent architecture, Comma, claiming to implement the "single Router + multiple Workers" multi-agent architecture proposed by DeepSeek Harness. The architecture aims to overcome the context bottleneck of a single session by introducing continuous learning, long-term memory, and 24/7 online capabilities, enabling agents to handle massive parallel tasks. The system is publicly available and the code is open-sourced.

According to multiple sources, Comma's core design uses a single Router to orchestrate multiple Worker agents, replacing the traditional serial single-session interaction model. The team highlights that human attention can only track 3-5 active sessions simultaneously, and argues that agents should take over massive parallel tasks. They introduce continuous context and persistent memory mechanisms to allow agents to run online for extended periods. However, current public information is mostly conceptual, lacking specific architectural parameters, benchmarks, and reproducible results.

This open-source initiative echoes the Super-agent vision proposed by the DeepSeek Harness team, addressing the core pain point that human attention becomes the bottleneck in AI applications. If validated, the architecture could push multi-agent systems from experimentation to production, providing developers with a scalable parallel task processing solution and potentially reshaping the interaction paradigm of AI assistants in long-term online scenarios.