Current AI agents remain trapped in a single-stream format, forcing them to complete one task—reading, thinking, or acting—before starting the next. This sequential bottleneck limits responsiveness and efficiency, a core constraint since early models like ChatGPT. Researchers now propose a fundamental architectural shift to unlock parallel computation.

By instruction-tuning models on multiple, simultaneous streams for reading, thinking, and output, each forward pass can process inputs and generate tokens across all streams at once. This data-driven change creates a causally dependent parallel system, moving beyond simple message exchanges to integrated, concurrent operation.

The approach promises tangible gains: faster agent response through true multitasking, improved security via better separation of concerns, and enhanced monitorability as distinct streams provide clearer insight into the model's process. This redefines how language models handle complex, real-time tasks.