Researchers have proposed a new architecture for persistent language agents, called Pera (Perception-Centered Architecture for Persistent Agents), in a paper published on Hugging Face. The architecture is designed to enable agents to continuously adapt their service procedures by perceiving signals from tasks, context, and environmental changes.
Traditional language agent frameworks typically focus on solving user-specified, bounded tasks. However, there is a growing need for agents that can provide persistent assistance in long-lived settings where user needs, context, and service procedures evolve over time. Pera addresses this gap by organizing agents around perception and control components that continually monitor episodic task executions, internal context, and environmental changes.
These perceived signals are used to construct 'lifecycle tasks' that drive the ongoing operation and adaptation of the agent's service procedures. The authors use Pera to retrospectively organize recent work in the field, examine a detailed case study, and offer insights for building more capable persistent agents.
The paper draws an analogy to software engineering's evolution from 'programming in the small' to 'programming in the large,' framing the development of language agents as a similar architectural transition toward long-lived, adaptive intelligence systems. The authors argue that Pera provides a framework for this transition, enabling agents to perceive relevant changes and adapt continuously.